A note before we start: this issue comes out slightly late compared to the usual schedule. I write these posts in my free time, so while I try to keep it consistent, holiday periods can affect my writing frequency. I’m sure you’ll still find the insights interesting and I have a few more articles in store that I think you'll enjoy. Now let’s get to it!

1. Industry news

Data centres: grid liabilities and assets

In the last week of June, a heat wave brought extreme temperatures to parts of Europe and the US. In the U.S. several transmission operators, like PJM, NYISO, ISO-NE, and IESO beat last summer’s load peak and challenged or exceeded this summer’s forecast as well.

Reserve margins in PJM (Eastern interconnection) shrank to ~3 %, day-ahead prices in ISO-NE (New England) cleared at >$400/MWh, and Long Island’s real-time LMP briefly spiked above $9,000/MWh. Emergency DR, voluntary curtailments and behind-the-meter solar helped keep the lights on, but also exposed how little controllable load remains on call.

I recommend reading GridStatus’ account of the events if you’re interested in electricity market dynamics.

The first heat wave of 2025 highlighted the escalating, year-over-year stress that extreme weather events place on our grid.

Data center consumption is a main driver of the load growth that’s exacerbating grid congestion. One of the consequences of this is that data center energy optimization is one of the keys to solving the challenges that electricity networks are facing.

Emerald AI, a startup based in Washington D.C., is trying to tap into this: during a field trial in an Oracle Cloud data center outside Phoenix, they managed to reduce the power consumption of AI workloads running on a cluster of 256 NVIDIA GPUs by 25% over three hours during a grid stress event, all while preserving compute service quality.1

The company, which just announced a 24.5M$ seed round, accomplished this by dynamically managing the different types of jobs running in an AI data center. Less time-sensitive tasks, like model training or academic fine-tuning, can be slowed down or paused. In contrast, critical workloads like real-time inference queries for a popular service aren't delayed but are instead redirected to a data center in a region with less grid stress. The full technical report on Emerald AI’s field trial can be found here.

This is extremely relevant: a recent study from Duke University found that approximately 100 GW of new data center capacity could be connected to the grid by making data centres demand more flexible. The logic is that the electricity grid usually operates with plenty of headroom. Its capacity is really only tested during brief periods of extreme demand, like heatwaves or winter storms. Consequently, there's often sufficient capacity on the existing grid to connect new data centers, provided they are designed to be flexible and can reduce their power consumption during these high-demand periods.

What I’m thinking

Being able to actively shift compute based on grid conditions, electricity price, and clean‑energy availability is key to deploying AI sustainably. With AI becoming more intertwined with energy, grid-aware computing is one of the most exciting spaces to be in, and there’s a bunch of companies doing very interesting stuff. ElectricityMaps for example, just released a carbon-aware scheduler to optimally schedule compute jobs based on the grid’s carbon intensity.

At the residential level, demand-side flexibility is also attracting attention, with Apple Home presenting its new EnergyKit that will allow load scheduling according to prices, and Octopus Energy introducing its first vehicle-to-grid bundle in the UK. This will allow users to access a BYD Dolphin, an EV charger, and access to a smart tariff that offers free home charging for around 300 £/month.

Most companies in the energy space are increasingly aware that a kWh is not just a kWh: when that kWh is being consumed makes all the difference. This, which wasn't that relevant for the past century, will be one of the main trends of the energy sector for the next decades.

If you want to get a grasp on the impact and size of price volatility, I recommend looking at this Python tutorial I wrote a couple of months ago:

Updates on the Iberian blackout

Last month I wrote about the Iberian peninsula blackout and contrasted different theories. Since then, Spain’s official investigations have clarified that the 28 April Iberian blackout was above all a multifactorial over-voltage / reactive-power control failure, not primarily a frequency-inertia event.23

System oscillations in the preceding hours, including an atypical one traced to a specific installation in southwest Spain, pushed voltages higher just as the system entered the critical midday window with insufficient dynamic voltage control. The investigations also noted that one of ten programmed thermal units was declared unavailable on the same day, but was not replaced by Red Eléctrica.

Although the anti-renewable narrative highlighted how renewable-induced low system inertia was to blame for the blackout, the investigations showed that inertia on 28 April actually exceeded ENTSO-E guidance, and Madrid reiterates the blackout was not caused by high renewable penetration.

What I’m thinking

I don’t see how this is an isolated incident and I believe this should be a wake-up call for all system operators. No grid is immune from the changes that are coming, and if you don’t plan ahead, these kinds of events will just keep happening more and more.

Red Eléctrica is under a lot of stress right now. They’re an entity whose sole responsibility is to guarantee the functioning and stability of the network and they’re not being able to do that. We need to start digitalizing our grids now and get creative about solutions because the challenges of modern electricity networks are not going away by themselves.

What is the net climate impact of AI?

Google’s latest sustainability report showed how quickly AI’s physical footprint is overwhelming corporate climate goals: its emissions are up 51% versus 2019, electricity use jumped 27% year-on-year, and Scope 3 supply-chain impacts tied to data-centre build-out surged 22% in 20244. Even aggressive clean-power contracting can’t outrun physics when growth is exponential and next-gen low-carbon sources like advanced geothermal or small modular reactors remain behind schedule.

Meta also made headlines with Zuckerberg announcing the construction of Hyperion, a 5GW cluster that will be approximately the size of Manhattan.

I’ve worked for the past 8 years on implementing AI for climate-positive applications, and I often wonder about the balance between the negative and positive impacts of AI on climate. The Net Climate Impact of AI Report by Jennifer Turliuk’s provides a good account of the variables to consider: on the one hand AI’s climate harms (infrastructure, manufacturing, water, even fossil exploration enablement), on the other its potential benefits (grid optimisation, predictive maintenance).

Current literature pegs realistic AI-enabled emissions reduction at perhaps 1.5–4% by 2030. But given the pace of data centre deployment, and considering the impact this will have on clean generation targets, there’s a right to be at least worried about the overall climate impact of AI. The University of Oxford also published a study about this, empirically finding an inverted-U curve: early AI adoption raises energy use and CO₂, but net environmental gains emerge only beyond high spending thresholds (~US$220–580 per capita).

The International Energy Agency is also turning its interest towards AI (who isn’t?). They recently launched the Energy & AI Observatory, that aims to track energy required to power AI, while also documenting AI applications in the energy sector. Give it a look here.

What I’m thinking

Besides the sustainability concerns over AI, which I extensively wrote about in previous articles, the map above shows a growing compute inequality. Most large data centers, those able to train and host state-of-the-art AI models, are concentrated in the U.S., China, and the European Union, leaving several countries from the developing world completely cut out. Coincidentally, energy is one of the reasons for the AI compute divide: in order to have these large clusters operating, governments need to have invested heavily in their energy system, both in terms of generation, transmission, and distribution of electricity.

This seems to be yet another trend that, if left to market dynamics, will increase global inequality and the dependency of countries from Africa, South America, and Asia on global superpowers.

2. Scientific publications

Reinforcement Learning for grid-interactive electrification

If we’re serious about the energy transition, we will have to transform very complex systems such as transportation, the buildings sector, and the electricity network. Traditionally, it’s been very difficult to model these systems with rule-based methods or traditional machine learning, due to their complexity and high number of variables at play. Reinforcement Learning is now offering a promising alternative solution for taking optimal decisions in these complex environments.

This past month three papers about this came out:

1. “Scalable reinforcement learning for large-scale coordination of electric vehicles using graph neural networks” introduces EV-GNN, a graph neural-network solution that models the complex relationships between EVs, charging stations, and the grid as a graph, which is then used to train a deep reinforcement learning agent to optimize charging.

2. “A scalable demand-side energy management control strategy for large

   residential districts based on an attention-driven multi-agent DRL approach” proposes an attention-based multi-agent reinforcement learning algorithm where each building in a district acts as an agent. These agents learn to cooperate in managing their energy consumption, storage, and generation to achieve collective goals like reducing peak demand and minimizing energy costs.

3. RL2Grid: Benchmarking Reinforcement Learning in Power Grid Operations introduces a new platform for developing and testing RL agents for grid control. The platform aims to provide a standardized and realistic simulation environment for tasks like managing power flow and preventing blackouts.

What I’m thinking

Reinforcement Learning is behind many of the breakthroughs of modern Machine Learning, including the training of state-of-the-art LLMs. In the energy sector, real-world applications didn’t take off yet. This is likely because energy is a critical domain where we need more confidence and research before letting self-learning algorithms take the reins. However, the escalating challenges we're witnessing, like the grid stability problems I covered earlier, mean we are now at a point where we are essentially forced to explore these advanced solutions. Over the next few years I believe we’ll see more and more of these applications moving from research stage to industry.

The code for both EV-GNN and RL2Grid is open-source and available on Github, a great place to start if you want to make it your summer side project to learn how real-world RL agents in the energy sector are built.

Blending Machine Learning and physics in the buildings sector

Traditional black-box Machine Learning methods often struggle when applied blindly to the buildings sector: they don’t understand the underlying physical processes that govern building dynamics, they struggle when there’s not enough data, and they lack interpretability, making it difficult for building operators and energy managers to trust their predictions or understand why certain control decisions are recommended.

Three trends that I’ve been following closely and that promise to solve these issues of traditional methods are transfer learning, digital twins, and physics-informed models.

Three papers recently came out last month tackling these topics:

1. GenTL: A General Transfer Learning Model for Building Thermal

   Dynamics. In this paper, researchers built a 3-layer LSTM pretrained on simulations from 450 archetypal Central European single-family houses: a universal source that can be fine-tuned to new targets with only 5–15% of a year’s data. The initial tests from the paper showed an average 42.1% RMSE reduction relative to conventional single-source transfer learning and substantially lower error variance across 144 target buildings.

2. A digital twin platform for building performance monitoring and

   optimization: Performance simulation and case studies. Researchers at LNBL built a modular, open, web-based digital twin platform that fuses live sensor/meter streams, external weather feeds, EnergyPlus physics models, and DER Functional Mock-up Units through an interoperable data pipeline. The platform can be used to simulate how the building would behave in different scenarios.

3. Integrating symbolic neural networks with building physics: A study

   and proposal. The authors of this paper explored physics-informed neural modeling and demonstrated that Kolmogorov–Arnold Networks (KANs) can rediscover fundamental equations for heat transfer, approximate complex formulas, and capture time-dependent dynamics. They also provide a handy flowchart to understand when KANs should be used vs other ANNs or traditional physical modeling.

What I’m thinking

At Ento we use a lot of traditional black-box Machine Learning to model the energy consumption of buildings. They’re really great for some of our use-cases, but in situations when:

1. Indoor sensors data is available

2. There’s gaps in the data

3. We want to actively control a building’s HVAC system

… plain black-box modeling often falls short. That’s why I’ve been closely monitoring these trends and I’m excited by the possibilities opened up by digital twin platforms and blending data-driven methods and physical modeling.

If you’re interested in an introduction to trade-offs and characteristics of different modeling strategies in the buildings sector, I wrote a detailed overview in a previous article:

3. Reimagine Energy content

I had fun presenting AI applications in the energy sector at the iEner‘25 conference in Madrid, organized by the Association of Energy Engineers.

Find the slides from the event below (in Spanish):

Energy optimization and verification using AI

3.48MB ∙ PDF file

Download

Download

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Senior Control and Optimization Engineer at EnergyHub

2. Power Systems Controls Engineer, Megapack at Tesla

3. Senior Data Science & AI Specialist at ABB

4. PhD in Machine Learning Techniques Applied To Power Grids at Escuela Politécnica Superior de la Universidad de Alcalá

   Support free content about AI in the energy sector by sharing this with someone in your network.

   Share

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

1. Industry news

How much energy does AI use?

This is a million-dollar question, and one that’s surprisingly difficult to answer. Reporters from MIT Technology Review published a great piece on this recently. I suggest to at least skim through it. It’s a very good introduction to the topic with several intuitive, but at the same time thorough, explanations. The authors measured the GPU energy consumption used for inference by different open‑source models. These were the results:

• 6700 J (1.8 Wh) for a text response with Llama 3.1 405B

• 4400 J (1.2 Wh) for an image generated with Stable Diffusion 3 Medium

• 3.4 MJ (0.9 kWh) for a 5-second video at 16 frames per second using CogVideoX

We don’t have equivalent data for flagship proprietary models unless their owners decide to publish it, but these open‑source benchmarks at least give us an order‑of‑magnitude sense of where today’s AI workloads sit on the energy scale.

If you want to get an intuition of this, you can try Chat UI, developed by researchers at Hugging Face, where you can see live GPU consumption estimation for different open-source models.

It’s also a great example of the LLM hallucination problem, as you can see in the image below.

If we extrapolate these numbers to multiple billions of daily queries (in December, OpenAI said that ChatGPT receives 1 billion messages every day1) we get to GWh consumption values per day. This does not even take into account the computation used for reasoning models. If we stop for a second and think how we will be soon surrounded by agents continuously running multiple tasks in the background for us, the total energy demand for AI reaches incredibly high numbers.

How much does this amount to today?

A few months back, we analyzed the 2024 report on data center energy usage published by Lawrence Berkeley Lab. The report says that data centers in the US used somewhere around 200 terawatt-hours of electricity in 2024, roughly what it takes to power Thailand for a year. AI-specific servers in these data centers are estimated to have used between 53 and 76 terawatt-hours of electricity. On the high end, this is enough to power more than 7.2 million US homes for a year.

What does the infrastructure required to support this look like?

As I’ve mentioned before, I really love a good map. And this one from NREL is worth more than a thousand words. Note how the average data center size increases considerably from existing facilities (yellow), to those under construction (orange), to proposed ones (white), with the latest reaching 2.4 GW, close to the entire generation capacity of Iceland2.

Beyond electricity consumption, what about water use?

Aside from the energy requirements, data centers also need significant water for cooling their circuits. Most large data-centers cool their servers with evaporative systems, meaning that, while water is recirculated, in each loop some water is lost to evaporation. To get an idea of the ballpark numbers, Google’s data centers consumed around 23B liters3 (6B gallons) of water in 2024, that’s just about the water used in a year by the country of Cabo Verde4. A recent investigation from Bloomberg News found that about two-thirds of new data centers built or in development since 2022 are in areas experiencing high levels of water stress5. And the trend seems to continue, with OpenAI and the U.S. government recently signing a deal to build massive data centers in the United Arab Emirates, starting with an AI campus in Abu Dhabi6. How does Abu Dhabi fare in terms of fresh water supply? Not well, as you might imagine.

Seawater cooling might sound like an obvious solution, but the salt quickly ruins ordinary pipes, meaning that pricey corrosion-proof infrastructure is needed. The raw intake also comes packed with sand, algae and marine critters that need to be filtered.

Powering the next decade of AI

When we look at data center consumption trends, the numbers are enough to scare any system operator. We might reach the 1,000 TWh of data centre consumption by 20307. McKinsey recently said that the global cost for scaling up AI infrastructure might be as high as $8 trillion, with more than $500 billion dedicated to power generation and transmission alone8.

Even knowing the price tag, the path to get there is anything but clear. Tens of GW have to be deployed each year, while also upgrading the grid infrastructure to be able to handle increased demand levels. Can this practically be achieved?

Nuclear Renaissance

Nuclear energy has been a very hot topic for the last few months: it’s carbon-free, it’s reliable, and, unlike renewables, it has support across the political spectrum. The Trump administration just signed four executive orders aimed at quadrupling America’s nuclear capacity over the next 25 years, bringing it from around 100 GW today to 400 GW by 20509. There is no clear path on how to do that, although the executive order mentions that 10 new reactors with completed designs must be under construction by 203010. The private sector is also gaining momentum: NuScale Power just won US approval for small nuclear reactors design, and Google announced a partnership with Elementl Power for three nuclear projects that will each produce 600 MW.11

While nuclear will certainly play an increasingly relevant role in the long run, it’s clear that the timelines don’t match. The AI compute increase is happening today, while these nuclear projects will take years, if not decades, to be operational.

Power companies are increasingly aware that, to sign large PPA contracts with big tech firms, the only viable way to scale production quickly is through a combination of renewables and natural gas. This is clear if we look at some of the latest deals in the space:

Renewables

1. Brookfield and Microsoft partner for 10.5 GW of renewables (May 2024)

2. Google, Intersect Power to develop co-located energy parks with $20B of renewables + storage (Dec 2024)

3. Amazon signs agreement to buy 476 MW of renewable power from Iberdrola in Spain (Feb 2025)

4. AES and Meta Sign Long-Term PPAs to Deliver 650 MW of Solar Capacity in Texas and Kansas (May 2025)

Natural gas

1. Constellation acquires Calpine’s gas fleet (27 GW) for $16B (Jan 2025)

2. Vistra acquires 2.6 GW of natural gas generation assets for $2B (May 2025)

3. NRG Energy Buys 18 Natural-Gas Power Plants (25GW), in a deal that includes a Virtual Power Plant platform (May 2025)

There have been two large acquisitions of existing natural gas assets in May, but in order to supply the incoming electricity demand increase, it’s clear that new gas-fired power plants will have to be built.

We can expect building new assets to cost more than acquiring existing ones, but how much more? Halcyon, a San Francisco-based company building an AI-assisted search and information platform for the energy sector, just announced their latest product, a US gas power plant tracker, which includes insights like the one below, which I found really interesting.

The acquisition prices outlined before are between 500$/kW and 1000 $/kW, buying existing assets costs then between one-fourth and one-half of what it costs to building them from scratch. But buying existing assets can only temporarily meet demand, and with solar's intermittency and regulatory challenges facing wind, it looks like we’ll see more and more gas-fired power plant buildouts over the next few months and years.

Geopolitical stakes

The interconnection of AI and energy has also implications on the geopolitical level. There is a global race towards achieving superintelligent AI. Both the U.S. and China understand that having ample energy for computing will grant a strategic edge in developing advanced AI, and potentially shift the global balance of power in their favor. OpenAI and other AI giants are using this as leverage to pledge faster data center build-out and more investment in grid infrastructure and energy supply. “Eventually the abundance of AI will come down to the abundance of energy, and in terms of long-term strategic investment I can’t see anything more important than energy” Sam Altman told a Senate committee, on May 8.12

At the same time, the U.S. has been forging AI-Energy alliances with oil-rich partner countries: a series of deals with Gulf states (Saudi Arabia, UAE, Qatar) aim to channel petrodollars and energy into building AI supercomputing centers aligned with U.S. interests.

While China seems to be slightly behind on chips, they also have a substantial lead on the clean energy generation side. According to a recent Carbon Brief analysis, China is set to produce over 600 TWh of electricity from newly added clean energy sources in 2025. For reference, the total US renewable generation capacity, including hydro and nuclear, is about 500 GW, and China added 375 GW of new capacity in 2024 alone, 75% of the US total clean generation capacity.

What I’m thinking

The race for global AI supremacy is on, and this will have profound implications on the energy sector over the next decade. With billions of dollars and geopolitical dominance on the line, companies are already turning to illegal practices to power their clusters13, rethinking their climate pledges14, and relaxing their safety policies15.

It looks like decarbonization and climate are now being sidelined. I expect this trend to only increase over the next few years. I mentioned this before and stand by it now, the only way in which we can effectively reduce emissions, in a world where abundant energy means having an edge in a global superpower conflict, is by making clean energy technologies make economic sense. I don’t see a future where we “voluntarily” reduce emissions while the geopolitical stakes are so high.

2. Scientific publications

Enabling flexible consumption in schools with Model Predictive Control

Field implementation of model-based predictive control in an all-electric school building: Impact of occupancy on energy flexibility: This study presents a field implementation of a grid-interactive model-based predictive control (MPC) system in an elementary school in Montreal, Canada. Researchers compared eight classrooms during cold winter days: four using a standard PI controller and four using the MPC.

During morning peak hours, when electricity is most expensive, the classrooms controlled by the MPC reduced their average power consumption by 61% compared to the classrooms running on a standard controller.

What I’m thinking

Schools are generally a great place to implement smart control algorithms, due to their very predictable occupancy schedules. At Ento, it’s been also been one our main targets for our control product.

Besides the standard heating output optimization based on weather and occupancy, I found it interesting that the authors were able to take into account the implicit demand response signal embedded in the utility’s electricity rates. With electricity rates becoming increasingly more volatile due to renewable penetration, having buildings that are able to intelligently adjust their consumption based on those signals is extremely valuable, and should eventually become a standard outside of research projects.

New techniques to slash the computational cost of LLM reasoning

Two papers from researchers at Salesforce explore a pragmatic solution to the high computational cost of complex reasoning. While Chain-of-Thought (CoT) prompting significantly improves performance on difficult tasks, the "thinking" steps make it slow and expensive. These studies show that you don't always need the full reasoning path to get the right answer.

The first paper, Fractured Chain-of-Thought Reasoning, introduces an inference-time strategy called Fractured Sampling. It breaks down reasoning into three controllable dimensions: the number of reasoning paths, the number of final answers per path, and the depth at which the reasoning is "fractured," or truncated. Through experiments on complex math and logic benchmarks, the authors demonstrate that exploring multiple partial reasoning paths is far more cost-effective than generating a few complete ones.

The second paper, Scalable Chain of Thoughts via Elastic Reasoning, proposes a training methodology to make models robust to this truncation. Elastic Reasoning explicitly separates the model's output into a "thinking" phase and a "solution" phase, each with its own token budget. Using reinforcement learning, they train the model to generate high-quality solutions even from incomplete thoughts.

What I’m thinking

Most of the energy consumption of LLMs comes from the inference phase, and CoT models have radically increased those costs. Any technique able to reduce computation while keeping quality to high standards should be explored. If we consider the scale at which LLMs are going to be used, such radical efficiency gains on every request would be more impactful than any power generation project that can be implemented. Coming from the building energy efficiency field, I see a very clear analogy: clean energy generation is important, but anything that can reduce energy use by increasing efficiency should be prioritized.

Code is available on Github and Hugging Face if you want to try it out.

3. Reimagine Energy publications

No full-length article this last month, but you can check out this LinkedIn post. It provides a good overview of the Python tutorials I released until now for solar production analysis.

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Staff Machine Learning Engineer at Sunrun

2. Data Engineer - Platform at Flower

3. Cloud Engineer (GCP) at GridEdge

4. Summer Internship - AI-Powered Climate Tech at TetraxAI

5. Postdoc position in energy & AI at KU Leuven

6. Postdoctoral Research Fellow (Multi-Scale Energy Systems) at NUS

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

1. Industry news

On April 28, a massive power blackout struck the entire Iberian Peninsula and parts of southern France. Around 12:33 CEST that day, lights went out for tens of millions of people in an unprecedented grid collapse for a European country. Essential services were disrupted: airports, hospitals, public transport, telecommunications, and industry all ground to a halt. Power restoration took many hours, with some areas (including my house in Barcelona) in the dark for more than eight hours before supply was fully reestablished.

While we’re still waiting for the official investigation to be concluded, and a report to be published, there are already interesting data that we can analyse. In this issue, I will unpack the facts known so far, what are the leading theories, and, of course, the role that AI can play in preventing future events like these.

Facts: what we know until now

Piecing together reports from ENTSO-E1, Spain's Red Eléctrica2, and various energy analysts3, the following timeline has been reconstructed:

1. In the hours and even days leading up to the main event, the Iberian and wider European grid experienced notable inter-area oscillations and voltage disturbances. On April 28, around 12:20 PM CEST, significant oscillations were observed, potentially stressing the system.

2. At 12:32:57 PM, in a rapid-fire sequence lasting about 20 seconds, at least three different generators tripped offline in southwest Spain, totaling around 2.2 GW. The reason behind these trips is still unknown.

3. This rapid 2.2 GW loss triggered a sharp decline in the Iberian grid's frequency. Protective measures, including Under-Frequency Load Shedding (UFLS), reportedly activated around 48 Hz. However, by 12:33:20 CEST, the phase angle difference between Iberia and the main European grid became too large, leading to the tripping of AC interconnectors with France. Iberia became an electrical island.

4. Once islanded and already severely generation-deficient, the Iberian system couldn't stabilize. A "massive disconnection of generation" occurred within the newly formed island, leading to a full blackout by approximately 12:33:24 CEST.

In the image below, we can see the three generation loss events within a few seconds, the frequency dropping in Málaga, as well as the rate of change of frequency (RoCoF), and the phase angle between Spain and continental Europe quickly increasing well over the 90° stability limit for transmission lines.

What does it mean that the phase angle difference led to the interconnectors with France tripping?

In an AC power system, voltage and current are not constant; they oscillate as sine waves. Generators across interconnected grids like Europe's must maintain synchronism, meaning aligned voltage frequencies and phase angles. When Spain lost significant generation, its generators slowed, causing its voltage phase angle to lag behind the stable European grid (represented by France). This growing phase difference triggered a massive power surge from France to Spain as the system tried to pull Spain back into sync. However, interconnectors have phase angle stability limits and exceeding these risks a dangerous loss of synchronism and widespread damage. To prevent this, protective relays detected the excessive phase angle difference and automatically tripped the interconnectors, safely islanding the Iberian Peninsula.

Renewables as the scapegoat

Many initial analyses of this event focused on the role of low system inertia in the blackout. Inertia, traditionally provided by the spinning mass of large fossil fuel and nuclear generators, acts like a shock absorber for the grid, resisting rapid changes in frequency when generation and load become imbalanced.

As grids integrate higher shares of inverter-based renewables (like solar and wind), which don't inherently provide this spinning inertia, the overall system inertia can decrease. The supporters of this maintain that had there been more inertia, the frequency wouldn't have dropped so quickly after the initial generation losses, giving the system operator more time to act and potentially averting the full collapse. This narrative leads to calls for slowing down renewable deployment or keeping fossil fuel plants online purely for their inertial properties.

An alternative explanation: overvoltages and Rate of Change of Frequency (RoCoF) relays

While low inertia undeniably makes a grid more sensitive to frequency deviations, there are also other theories. Luis Badesa, a Professor of Power Systems at UPM, points out that the initial 2.2 GW loss constituted an N-3 contingency (at least three separate generation losses). He then argues that no Transmission System Operator (TSO) would typically plan or carry enough inertia to withstand an N-3 event, especially while subsequently operating as an electrical island. The real problem, therefore, started before the final, inertia-exacerbated frequency collapse.

The key questions then become:

• What caused the initial, correlated generator trips? There has been no official confirmation yet, but different analysts and reports reports suggest overvoltages in Southwest Spain, possibly linked to the earlier inter-area oscillations, could be a strong candidate. If so, the initial problem was one of voltage stability and a cascade of trips triggered by that, rather than a primary frequency issue.

• What caused the final "massive disconnection of generation" after Iberia was islanded? While we’re still waiting for Red Eléctrica and ENTSO-E to provide an official report, the hypothesis advanced by Prof. Badesa is that the N-3 event caused the grid frequency to change really quickly, potentially triggering Rate of Change of Frequency (RoCoF) protection relays for more generators. RoCoF relays are protective devices that monitor the speed of frequency change in a power system. By detecting the rapid frequency change that often occurs when grid connection is lost, the RoCoF relay trips the generator before it can operate in an unstable islanded mode for a prolonged period. This prevents potential damage from off-nominal frequency/voltage operation. This is how the N-3 event could have cascaded into a complete system shutdown.

According to this theory, low inertia played a role in the final stage of the collapse within the island, but it wasn't necessarily the root cause of the initial N-3 contingency that pushed the system into a highly unstable condition.

Enhancing grid resilience and demand-side flexibility

If Professor Badesa's line of reasoning holds, then the most effective solutions go beyond simply adding inertia or relying more on fossil fuels. Instead, we need to focus on:

• Preventing the N-3 event: This means thoroughly investigating and addressing the root cause of those initial correlated generator trips. If overvoltages were the issue, there are solutions such as enhanced voltage control, better damping of inter-area oscillations, robust network infrastructure, and appropriate generator protection settings for voltage ride-through.

• Revising RoCoF protection strategies: RoCoF protective relays shouldn't trip prematurely before faster-acting resources can respond, especially in an islanded, low-inertia scenario.

• Deploying rapid, large-scale Demand-Side Flexibility and storage: Had the Iberian island been able to rapidly shed over 1-2 GW of load (DSF) or inject a similar amount of power (from batteries providing Fast Frequency Response) within seconds of the 2.2 GW loss, it might have stabilized frequency quickly enough to prevent widespread RoCoF trips and the final collapse.

• Installing Grid-Forming Inverters (GFMs): GFMs are inverters that can rapidly inject or absorb power in response to frequency deviations, mimicking the stabilizing inertial response of spinning generators. Integrating renewables and batteries with GFM capabilities would allow them to actively contribute to grid stability by providing synthetic inertia and maintaining voltage and frequency.

• Improving interconnection with France: As visible in the image below, the interconnection between Spain and France is limited to a few points across the Pyrenees. This is a known bottleneck that restricts the amount of power that can be exchanged between the Iberian Peninsula and the wider Continental European grid. While the interconnectors correctly tripped to protect the larger system during this event, greater interconnection capacity would offer several benefits for resilience.

What I’m thinking

Many analyses blaming renewables come from authors without power‑system expertise and often lack data. These interpretations are also highly politicized, and come from organizations attempting to advance their own agendas.

Without more information, we can’t draw definitive conclusions at this stage. All I will say is that if, as has been suggested, the incident involved three generators tripping due to overvoltages, then, while increased system inertia is certainly beneficial for the resilience of any power system, it likely could not have prevented this blackout.

Regardless of this event’s specific cause, several concurrent trends are undermining the stability of our grids: higher renewable penetration, the electrification of heat and mobility, the growth of data centres, and more frequent extreme weather events. While no single technical solution exists for this massive engineering challenge, two pillars will certainly be the digitalization of the grid and enhanced control of demand-side resources in response to grid conditions. There were 19 seconds between the first and second generation loss events. This timeframe could have been sufficient to disconnect a fleet of EV chargers, reduce the load of a data centre, or temporarily curtail HVAC systems in a portfolio of commercial buildings. This might have mitigated the impact on frequency, potentially avoiding the subsequent islanding and shutdown. With the economic damage from this blackout estimated at 400 million euros4, this appears as an extremely complex problem, but one that is clearly worth solving.

Have you explored any interesting theories behind the Iberian blackout? I would love to hear them!

Leave a comment

2. Scientific publications

Causally predicting control scenarios with Double Machine Learning

Beyond correlation: A causality-driven model for indoor temperature control introduces a double‑machine‑learning (DML) framework that learns the causal impact of heat-pump set‑point temperature on indoor air temperature. The algorithm was tested in a 38 m² office lab using 3,781 ten‑minute datapoints collected over 28 days.

What I’m thinking

Standard machine‑learning models often rely on statistical correlations that don’t reflect true cause‑and‑effect, leading to poor performance under new conditions. When directly controlling buildings, this is usually solved by using grey‑box models that blend physical and statistical modeling.

The proposed DML approach seems interesting in that it allows us to define a causal “what‑if” scenario. This could prove very useful when assessing the potential impact of different control strategies and evaluating the trade‑off between energy savings and comfort. One issue I see is that it was tested in a small office space with a lot of monitoring. Translating this to larger buildings with less data will certainly be a challenging task.

A new building dataset with 10,000 timeseries data points

BTS: Building Timeseries Dataset: Empowering Large-Scale Building Analytics. This paper, presented at NeurIPS 2024, introduces a dataset that covers three diverse Australian buildings over a three-year period, encompassing over 10,000 timeseries data points and 240 unique ontologies, all standardized using the Brick schema. The data and accompanying code are published openly on Github.

What I’m thinking

It’s always exciting when datasets with so much detail are published. By keeping the data raw, including missing values, mislabeled alarms, etc., the authors give us something closer to production reality than to a Kaggle competition. This can be great for stress‑testing automated cleaning, anomaly‑detection workflows, and defining control strategies that are robust to real-world operation.

3. Reimagine Energy publications

In my latest Python tutorial I use publicly available data to show how rooftop solar PV savings are affected by volatile electricity markets in Germany.

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. PhD student in development of digital twins for buildings at Uppsala University

2. Data and AI Strategist at Siemens

3. AI Engineer - Energy and Utilities at CGI

4. Data Science Forecaster at Eneco

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

I collaborated with Julien on the ideas presented in this article, and he has generously provided an introductory commentary:

The rise of solar capacity has been arguably the largest game changer in the electricity sector of recent years. This has been made possible by the fast declining cost of solar panels, as well as the extreme modularity of solar projects (from less than 1 kW to several hundreds of MW) and the possibility of fast deployment.

As solar has an obvious daily and seasonal pattern, electricity markets have been affected accordingly: abundant energy at midday from March to November causing prices to crash. This has led to the so-called solar cannibalization effect which has been accelerating since the end of the energy crisis. It is becoming increasingly crucial to consider this dynamic if we want to keep the decabonization of the power sector a reality.

Benedetto's post contributes to the knowledge diffusion of this important matter.

— Julien Jomaux

This analysis builds on the series of Python tutorials I’ve published on simulating rooftop PV production data and matching it with building load profiles. You can find the complete list of my Python tutorials here. In the examples that follow, I’ll assume you’re already familiar with:

• Loading a building’s hourly electricity consumption data (in the variable hourly_site_consumption)1

• Generating the hourly electricity production of a PV system using pvlib (in the variable module_energy)2

• Analyzing these time series to see how much electricity is consumed on site vs. injected back into the grid3

For this post, I imagine an alternative scenario where our previously analyzed U.S. building is located in Germany—purely to work with German electricity price data.

As a first step, I downloaded hourly pricing data from the Ember platform. The data is sourced and cleaned from ENTSO-E, the European Network of Transmission System Operators for Electricity.

Here’s how the data is loaded and prepared:

import pandas as pd

electricity_price = pd.read_csv('data/germany-electricity-prices.csv')
# Convert datetime columns to datetime type
electricity_price['Datetime (Local)'] = pd.to_datetime(electricity_price['Datetime (Local)'])
# Extract only prices from 2024
electricity_price = electricity_price[electricity_price['Datetime (Local)'].dt.year == 2024]
# set datetime as index
electricity_price.set_index('Datetime (Local)', inplace=True)
# Get the price column and convert from EUR/MWh to EUR/kWh
electricity_price = electricity_price['Price (EUR/MWhe)'] / 1000
avg_electricity_price = electricity_price.mean()

Let’s start by visualizing the average daily price profile for each month of 2024. This will give us a quick look at how prices vary across hours:

import plotly.graph_objects as go
# Create a daily electricity price profile for the whole year
electricity_price_with_hour = electricity_price.copy()

# Convert to DataFrame if it's a Series
electricity_price_with_hour = electricity_price_with_hour.to_frame()

# The index is already a DatetimeIndex, so we can directly extract components
electricity_price_with_hour['hour'] = electricity_price_with_hour.index.hour
electricity_price_with_hour['date'] = electricity_price_with_hour.index.normalize().date

# Create a figure for monthly average electricity price profiles
fig = go.Figure()

# Extract month from the datetime index and create a new column
electricity_price_with_hour['month'] = electricity_price_with_hour.index.month

# Calculate the average price for each hour within each month
monthly_avg = electricity_price_with_hour.groupby(['month', 'hour'])[electricity_price.name].mean().reset_index()

# Get unique months
unique_months = sorted(monthly_avg['month'].unique())

# Add a trace for each month
import calendar
for month in unique_months:
    month_data = monthly_avg[monthly_avg['month'] == month]
    month_name = calendar.month_name[month]
    fig.add_trace(go.Scatter(
        x=month_data['hour'], 
        y=month_data[electricity_price.name],
        mode='lines',
        name=month_name,
        line=dict(width=2)
    ))

fig.update_layout(
    title='Average Monthly Electricity Price Profiles',
    xaxis_title='Hour of Day',
    yaxis_title='Average Electricity Price (EUR/MWhe)',
    xaxis=dict(tickmode='linear', tick0=0, dtick=1),
    legend_title='Month'
)

fig.show()

Output

From this plot, we see a characteristic “duck curve”, where prices tend to dip during peak solar generation. This reflects Germany’s substantial solar penetration: as more PV capacity comes online, midday prices often drop (or even become negative), reducing the economic returns for newly added solar capacity.

Next, let’s combine this pricing data with our building’s load profile and a simulated rooftop PV system. We’ll assume:

• A 1.7 MWp rooftop system (system_size = 1 700 kWp), consistent with some of my previous examples, where we found 1.7 MWp to be near the “optimal” size for this particular building.

• An hourly production time series called module_energy (kWh per panel per hour, from pvlib). In order to estimate this, you’ll need to use Germany’s weather data, which you can find on the German Meteorological Service website.

• A load profile hourly_site_consumption (kWh per hour).

With these values, we can calculate onsite consumption, grid consumption, and grid injection:

system_size = 1700 #kWp
module_rated_power = 0.4 # kWp
panel_count = system_size / module_rated_power
# match the PV production to the consumption
pv_production = module_energy * panel_count / 1000  # convert from Wh to kWh
# calculate the grid consumption as the difference between the consumption and the production (but not less than 0)
grid_consumption = (hourly_site_consumption - pv_production).clip(lower=0)
# calculate self consumption (electricity that is consumed on site)
self_consumption = hourly_site_consumption - grid_consumption
# calculate grid injection (excess electricity injected into the grid)
grid_injection = pv_production - self_consumption

By matching the electricity flows and prices, we can now estimate the annual ROI of this PV system. The idea is to compare the return obtained using the average price, versus the actual hourly electricity prices. For this analysis, we’ll only consider the return coming from self-consumption, as the regulatory framework to receive compensation for injecting electricity into the grid with large PV systems is quite complex in Germany.

Let’s first estimate the annual economic returns considering the average electricity price:

total_return_fixed_price = self_consumption.sum() * avg_electricity_price

print(f'The total annual return, considering an average electricity price of {round(avg_electricity_price, 3)} EUR/kWh, is: {round(total_return_fixed_price)} EUR')

Now let’s calculate the annual return using the actual hourly price whenever the building is self-consuming electricity:

total_return_hourly_price = (self_consumption * electricity_price).sum()
print(f'The total annual return, considering the dynamic hourly electricity price, is: {round(total_return_hourly_price)} EUR')

Finally, compare the difference:

print(f"The annual return difference between the fixed and dynamic price is: {round(total_return_hourly_price - total_return_fixed_price)} EUR, {round((total_return_hourly_price - total_return_fixed_price) / total_return_fixed_price * 100, 2)}% of the fixed price return")

Output

In the example run I performed, the system’s annual income was around 35% lower when considering real-time prices instead of a constant average. In monetary terms, that amounted to about 30k EUR less than expected. The reason is straightforward: the hours with the highest PV output often coincide with the lowest (or even negative) electricity prices, eroding the financial benefit of self-consumption.

Implications and Possible Solutions

So, are there any solutions to this, or is solar doomed as a technology?

The first and most obvious solution is energy storage: with battery costs decreasing and price volatility growing, coupling rooftop PV with battery storage becomes increasingly attractive. By shifting excess generation to higher-price hours, storage can recapture some of the lost economic value.

This LinkedIn post I saw a few days ago summarizes this quite well:

The other solution, which might have an even bigger scale given that it’s mostly based on software rather than hardware, is Demand-Side Flexibility: adjusting building operations to consume more energy during hours of high renewable generation.

To exemplify this last point, for the building we’re analyzing, let’s have a look at how the load might align with PV generation and hourly prices during a sample week in August, when the “duck curve” is quite pronounced.

First, let’s look at electricity prices and solar production:

# Compare load, generation and price for one week in August
august_start_date = pd.Timestamp('2024-08-01').date()
august_end_date = pd.Timestamp('2024-08-08').date()

# Filter data for the selected week
# Make sure all dataframes have the same index length
common_index = hourly_site_consumption.index.intersection(pv_production.index).intersection(electricity_price.index)

# Create mask for August week
august_mask = [(idx.date() >= august_start_date) & (idx.date() < august_end_date) for idx in common_index]
august_indices = common_index[august_mask]

# Create figure with secondary y-axis for August
from plotly.subplots import make_subplots

fig = make_subplots(specs=[[{"secondary_y": True}]])

# Add traces for load and generation on primary y-axis
fig.add_trace(
    go.Scatter(
        x=august_indices, 
        y=pv_production.loc[august_indices], 
        name='Solar Generation (kW)',
    ),
    secondary_y=False, 
)

# Add trace for electricity price on secondary y-axis
fig.add_trace(
    go.Scatter(
        x=august_indices, 
        y=electricity_price.loc[august_indices], 
        name='Electricity Price ($/kWh)',
    ),
    secondary_y=True
)

fig.show()

Output

We can see how the two time series are almost perfectly inversely correlated. Whenever the sun is shining, the price is lowest. This means that there are some hours when this building could use loads of electricity that would either be offset by high on-site generation or purchased at very low cost.

Let’s have a look at how the building load relates to this:

# add electricity consumption timeseries
fig.add_trace(
    go.Scatter(x=august_indices, y=hourly_site_consumption.loc[august_indices], name='Building Load (kW)'),
    secondary_y=False
)
fig.show()

Output

We can see that most building operations already take place during hours of high PV production, but there’s often room to fine-tune controls (while maintaining occupant comfort) to capture additional economic benefit. This could mean ramping up building operations slightly later, or even temporarily shutting down some of the equipment during the day, depending on the solar production and electricity price signals.

Heat maps are also really powerful at outlining this:

# generate heat maps for solar production, price, and consumption
# First, ensure all data series have the same length by aligning them
common_index = pv_production.index.intersection(electricity_price.index).intersection(electricity_price_with_hour.index)

# Reindex all series to the common index
pv_production_aligned = pv_production.loc[common_index]
electricity_price_aligned = electricity_price.loc[common_index]
electricity_price_with_hour_aligned = electricity_price_with_hour.loc[common_index]
hourly_site_consumption_aligned = hourly_site_consumption.loc[common_index]

heatmap_data = pd.DataFrame({
    'Solar Production (kWh)': pv_production_aligned,
    'Electricity Price (EUR/kWh)': electricity_price_aligned,
    'Consumption (kWh)': hourly_site_consumption_aligned,
    'Hour of Day': electricity_price_with_hour_aligned.index.hour
})

# Extract hour and month from the datetime index
heatmap_data['Hour'] = heatmap_data.index.hour
heatmap_data['Month'] = heatmap_data.index.month

# Create month names for better readability
month_names = {
    1: 'January', 2: 'February', 3: 'March', 4: 'April', 
    5: 'May', 6: 'June', 7: 'July', 8: 'August', 
    9: 'September', 10: 'October', 11: 'November', 12: 'December'
}
heatmap_data['Month_Name'] = heatmap_data['Month'].map(month_names)

# 1. Electricity Price Heatmap
price_pivot = heatmap_data.pivot_table(
    values='Electricity Price (EUR/kWh)', 
    index='Month_Name', 
    columns='Hour',
    aggfunc='mean'
)

# Reorder months chronologically
price_pivot = price_pivot.reindex([month_names[i] for i in range(1, 13)])

# Create the price heatmap
fig_price_heatmap = px.imshow(
    price_pivot,
    labels=dict(x="Hour of Day", y="Month", color="Avg. Price ($/kWh)"),
    x=[str(h) for h in range(24)],
    y=price_pivot.index,
    color_continuous_scale="Viridis"
)

fig_price_heatmap.update_layout(
    width=900,
    height=600,
    xaxis=dict(tickmode='linear')
)

fig_price_heatmap.show()

# 2. Solar Generation Heatmap

solar_pivot = heatmap_data.pivot_table(
        values='Solar Production (kWh)', 
        index='Month_Name', 
        columns='Hour',
        aggfunc='mean'
    )
    
# Reorder months chronologically
solar_pivot = solar_pivot.reindex([month_names[i] for i in range(1, 13)])
    
# Create the solar generation heatmap
fig_solar_heatmap = px.imshow(
        solar_pivot,
        labels=dict(x="Hour of Day", y="Month", color="Avg. Solar Production (kWh)"),
        x=[str(h) for h in range(24)],
        y=solar_pivot.index,
        color_continuous_scale="YlOrRd"
    )
    
fig_solar_heatmap.update_layout(
        width=900,
        height=600,
        xaxis=dict(tickmode='linear')
    )
    
fig_solar_heatmap.show()

# 3. Building Consumption Heatmap

consumption_pivot = heatmap_data.pivot_table(
            values='Consumption (kWh)', 
            index='Month_Name', 
            columns='Hour',
            aggfunc='mean'
        )
        
# Reorder months chronologically
consumption_pivot = consumption_pivot.reindex([month_names[i] for i in range(1, 13)])
        
# Create the consumption heatmap
fig_consumption_heatmap = px.imshow(
            consumption_pivot,
            labels=dict(x="Hour of Day", y="Month", color="Consumption (kWh)"),
            x=[str(h) for h in range(24)],
            y=consumption_pivot.index,
            title="Average Building Consumption by Hour and Month",
            color_continuous_scale="Blues"
        )
        
fig_consumption_heatmap.update_layout(
            width=900,
            height=600,
            xaxis=dict(tickmode='linear')
        )
        
fig_consumption_heatmap.show()

Output

The lowest electricity prices and highest PV production align quite well.

The building load analysis, on the other hand, shows that the consumption is driven by summer cooling, as the hours of highest consumption are in the middle of the day during the summer. Ideally we’d want to align that heat map as much as possible with the electricity price and solar generation maps, to have the lowest possible impact.

Conclusion

The German economy heavily relies on industry, and a stable, affordable power supply is essential for these industries to flourish. It’s no surprise, then, that the impact of renewables on Germany’s electricity prices, including events like the ‘Dunkelflaute’, has dominated public discussion in recent months, becoming one of the critical factors leading to the government’s collapse in December.

However, before the government fell, it managed to pass an energy reform4 aimed at tackling these issues, notably by promoting energy storage co-located with renewable energy sources. Two key points from this reform are:

• No feed-in tariff will be paid to renewable energy producers during hours when electricity prices are negative.

• It is now allowed to charge energy storage systems co-located with renewable assets—a practice previously prohibited by the so-called “exclusivity principle”.

It’s increasingly clear that renewables without accompanying energy storage or demand-side flexibility can disrupt electricity markets and networks. This isn’t surprising, considering it has already been an issue in California for more than ten years.

I’m curious to see what new business models will emerge from these European market conditions, especially when combined with data, advanced metering infrastructure, and AI-driven solutions.

1. Industry news

A breakthrough in image generation

OpenAI released a new image generation model and announced gaining 1 million new users in one hour thanks to it. I wrote a thorough analysis of the models’ capabilities (and shortcomings) in my latest article. Give it a read if you haven’t already:

What I’m thinking

One of the most impressive features of the new OpenAI image generation model is its ability to accurately edit an image following any artistic style. As the internet quickly realized this, images generated with the signature aesthetic from Japanese film animation studio Studio Ghibli went viral. While the model’s ability to replicate Studio Ghibli’s style is impressive (and entertaining), I also have significant concerns:

1. The appropriation of vast amounts of artists' work scraped from the internet without consent, credit, or compensation by large AI corporations seems ethically dubious at best.

2. The beauty of something like the Ghibli aesthetic isn’t just about the visual output. It is the meticulous craftsmanship, the human dedication, the process and detail involved. German philosopher Walter Benjamin writes in his essay The Work of Art in the Age of Mechanical Reproduction about the “aura” of an artwork, and how that is linked to its uniqueness and authenticity. I fear that the ability to mass produce any artistic style, aesthetic or imagery will contribute to the de-evaluation of such pieces.

3. How much electricity for compute, and water for data center cooling, did it take to “Ghiblify” the internet? OpenAI doesn’t share any kind of data about this. We need to demand more transparency from LLM companies so that we can better decide which models to use and how frequently.

Gemini takes the lead

While OpenAI topped all the headlines because of its new image generation model, Google quietly took the lead in the “smartest model” race with Gemini 2.5.

If you’re curious about the latest updates on which models are best at what, this is a good overview: Claude is still the best at coding, but Gemini 2.5 now leads in almost all other areas. OpenAI o1 is still the best for multilingual tasks, closely followed by Llama 70B and Mistral.

What I’m thinking

If you haven’t tested a reasoning model yet, I suggest playing around with Gemini 2.5 on Google AI Studio (it’s free for now), to experience the current capabilities of best-in-class LLMs.

The API pricing for Gemini 2.5 is still unknown, but considering Gemini 2.0 has one of the lowest price/token rates (30 times lower than GPT-4o), there may soon be little reason to use OpenAI’s API unless they release a significantly better model.

Model Context Protocol: extending AI’s reach into the physical and operational world

One trending concept this past month was the Model Context Protocol (MCP), created and open-sourced by Anthropic late last year. MCP’s core purpose is to act as a standard communication layer between AI models and external data/tools. It defines how an AI agent can query or interact with external systems in a uniform way. A popular analogy is that MCP is like a “USB-C port” for AI: a single, standardized connector that works across many types of devices (in this case, data sources, tools, and applications).

In the buildings sector, MCP could serve as the connection between an AI assistant and a building’s automation system. By deploying MCP servers that interface with IoT devices, an AI agent can manage the environment through natural language commands or autonomous policies. EMQ, a software provider for open-source IoT data infrastructure, just published an overview of potential solutions to use MCP in combination with MQTT, another messaging protocol designed for machine-to-machine communication. One example from their blog would be a scenario in which a user could say, “I’ll be home in an hour, set the living room temperature to 25°C and humidity to 40%,” and behind the scenes an AI using MCP would communicate with the thermostat and humidifier to carry that out.

Home Assistant, a popular open-source home automation software, already added MCP support in their February update. If you're using it, you can test MCP now.

What I’m thinking

MCP extends AI’s reach into the physical and operational world, making it highly relevant for energy and buildings. In a demand-side flexibility scenario, for instance, an AI agent might react to certain signals by temporarily reducing HVAC usage during grid stress. These complex cross-system optimizations become more feasible when an AI can communicate with all necessary subsystems via one protocol.

One concern from my side: we’re increasingly enabling AI agents to operate autonomously in the world, which is good… until it isn’t. This doesn’t mean we shouldn't develop this technology, but we must concurrently research and invest in AI safety, build guardrails, and remain mindful of potential misuse.

2. Scientific publications

Tracing the thoughts of large language models

Two interesting research papers from Anthropic were released this past month. Auditing Language Models for Hidden Objectives, investigates the feasibility of detecting hidden undesirable goals within models. Researchers trained a model with a hidden objective and conducted blind auditing games to test detection methods. Results showed that teams using interpretability tools, behavioral attacks, and training data analysis could successfully uncover the hidden objective, while teams limited to black-box interaction (API access) could not. The second paper, On the Biology of a Large Language Model, explores the internal computational mechanisms of Claude Haiku, employing circuit tracing to reverse-engineer how the model performs tasks such as reasoning, planning, and multilingual processing. Intuitively, this was achieved by replacing the model's complex neurons with simpler, understandable "features" and mapping the flow of information between them.

What I’m thinking

The idea of LLMs being used by governments and corporations to control and influence how people think is deeply concerning. xAI’s Grok, for instance, has already been called out for sharing disinformation or censoring criticism of Elon Musk and Donald Trump. Within this context, being able to investigate potential biases maliciously introduced in LLMs seems very relevant.

Another interesting point is that the circuit tracing methods introduced in the second paper might be used to improve the audits discussed in the first. Instead of relying solely on finding suspicious training data or specific behavioral failures, auditors could potentially use circuit analysis to directly inspect for known problematic mechanisms or to understand the root cause of a behavior flagged during an audit.

While no company is “perfect”, Anthropic is among the few dedicating resources to interpretability and safety research, I hope they maintain this focus!

Generating synthetic fault data with GANs

Hybrid-CGAN: A Hybrid approach combining simulation and generative adversarial networks for fault detection and diagnosis in buildings. In this paper, authors from the National University of Singapore propose combining Generative Adversarial Networks (GANs) with simulation data from EnergyPlus to generate realistic synthetic fault data, reducing the need for actual, hard-to-get fault samples. The study focuses on faults within an Air Handling Unit (AHU) system, representative of a large office building. Results showed synthetic data generated in this way is significantly more realistic than simulation data alone.

What I’m thinking

Fault detection data is very hard to retrieve and typically locked in proprietary BMS systems. We previously discussed how data scarcity is a significant issue in the sector and how transfer learning is one of the potential solutions. Generating synthetic data is yet another solution to this problem and a promising one for certain applications. The authors mention that the data will be made available on request. It would be great if they could open-source both code and data, so that the community could validate and advance this very relevant research!

3. Reimagine Energy publications

I already mentioned at the beginning of this issue, this month’s post is an accessible read that you can share with all your energy-loving friends and colleagues:

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Data Scientist & Building Energy Specialist at Open Technologies

2. Lead Data Scientist for DER & Policy Forecasting at National Grid

3. AI Controls Engineer – HVAC Automation at Brainbox AI

4. PhD scholarship in Impact and Scalability of Advanced Controls for Energy Demand Flexibility in Buildings at DTU - Technical University of Denmark

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion.

ChatGPT gained 1 million new users in just 1 hour on Monday as the world became “Ghiblified”.

Setting aside the art plagiarism discussion, which I won’t address here, I believe this marks a breakthrough that will permanently change how humanity interacts with visual content. The ability to realistically edit existing images and iteratively improve results through a chat interface marks all of the difference with previous versions. Before last week, image generation with ChatGPT was very dull and unusable in any practical sense. Now it just entered the realm of tools that can be useful in everyday personal and professional settings.

Another exciting (though almost unnoticed) release last week was Google’s Gemini 2.5. This model currently tops all benchmarks and might very well be the most intelligent model available to date. It is temporarily accessible for free on Google AI Studio. I recommend you check it out.

Amidst all the hype, I thought it would be fun to test both models, so I asked Gemini 2.5 to generate a few use-cases to test GPT-4o image generation in the “buildings, energy, and environment” field. Some of them were not that interesting or relevant, among the rest, I decided to test the top five:

1. Renewable Energy Project Site Conceptualization: Generate realistic concept images for proposed renewable energy projects (wind, solar, storage) in specific environments, incorporating multiple elements accurately.

2. Illustrating Building Energy Efficiency Retrofit Scenarios: Upload an image of an existing building (house, office). Prompt GPT-4o to generate "after" images showing the visual impact of specific energy efficiency upgrades.

3. Generating Annotated Diagrams of Complex Energy Systems: Create clear, visually appealing diagrams explaining complex energy systems or processes for educational or communication purposes.

4. Visualizing Energy Audit Recommendations: After a conceptual energy audit, use GPT-4o to create visuals for the report.

5. Component Identification and Installation Guides: Upload a picture of an energy system component (e.g., a specific type of solar inverter, a smart meter). Ask GPT-4o to generate a simplified diagram identifying key ports or indicators with labels, or create a visual showing the basic steps for mounting it.

1. Renewable Energy Project Site Conceptualization

This use-case is especially interesting: a few months ago we tested the ability of Google's Solar API to detect the area of a rooftop suitable for solar panels installation. I decided to try this on same roof that we analysed in our former article: a building in Washington D.C.

Prompt: Edit this image highlighting the areas of the rooftop that are suitable for the installation of PV panels.

Output

Prompt: Can you show how the rooftop would look like with panels installed in those areas?

Output

The result is not great: the initial image was cropped, so some of the roof is not visible anymore. The highlighted areas in the first image are all on sections of the roof, but some areas are left out without an apparent reason. Most importantly, the panels in the second image do not correspond to the areas marked in the first image, so it looks like the model is not able to accurately detect the areas that it itself marked.

The result is not that different from the one we get using Google Solar, but I wouldn’t trust using this in a professional setting.

2. Illustrating Building Energy Efficiency Retrofit Scenarios

I thought it would be interesting to test this with an image from a real building. I have some old fixtures at home that are not that well insulated, so I decided to ask ChatGPT how a retrofit would look like.

Prompt: I'm writing a building energy retrofit proposal plan for a residential unit. Edit this image to show how it would look if triple glazing windows were installed.

Output

That’s quite good. I go on to ask for solar shading as well.

Prompt: Now add solar shading as well.

That’s not bad, although that’s not the type of shading that I was thinking about.

Prompt: Change the solar shading you added with an outside awning.

Pretty good, I also want to ask it to add some notes on the image showing the different elements that were added plus the estimated cost range.

Prompt: Now mark the improvements (triple glazing windows and solar shading) on the image with a colored pencil. Include the price range for both. Make the image brighter.

It didn’t nail the labels, although maybe they could be fixed with more specific prompting.

The final result in this case is quite impressive. I believe this could be potentially used in a professional setting.

3. Generating Annotated Diagrams of Complex Energy Systems

Here I wanted to test the creation of an image with educational purposes: a diagram showing energy flows for a house equipped with solar and batteries.

Prompt: Generate a diagram of a residential grid-tied solar PV system with battery storage. Include: solar panels, inverter, battery pack, smart meter, connection to the grid, and main electrical panel. Show arrows indicating energy flow during daytime charging and nighttime discharging.

That’s a decent start. If I wanted to use this in the newsletter, it would need to match my colors and branding though. Let's see if it can do this.

Prompt: Edit the image to use the colors of my brand.

Output

Now the colors are correct, but the flows are not. I might ask it to generate just the icons, and then draw the arrows and text on top myself, but then there’s not that much of an advantage over doing the design directly on Canva. My conclusion here is that the model is not yet able to understand the logic behind the image: the sun should be hitting the panels and not the inverter, the switchboard should be connected to the smart meter, etc. For this use-case, I don’t think we’re ready for professional or educational use.

4. Visualizing Energy Audit Recommendations

I decided to use here a typical suggestion that might be included in an energy performance certificate.

Prompt: I'm drafting an energy audit for a commercial building. After inspecting the building, I identified several measures that can be taken to improve the energy efficiency of the building. I will send you the description of one of them. Create an image that I can add to the audit report to better explain the proposed action.

EEM: Internal Insulation

Description: This measure involves installing internal thermal insulation, specifically using 9cm thick calcium silicate panels finished with plasterboard, on the walls of the office area. The goal is to improve the thermal performance of the building envelope and reduce heat loss.

Output

Prompt: Make it photorealistic.

Output

There are a few issues with the sizes and distances, but apart from that the result is quite good. I think this could be used for professional or educational use.

5. Component Identification and Installation Guides

For this use-case I sent the image of a common house boiler and asked it to identify all the components.

Prompt: Mark the main components of this boiler directly on the image.

Output

Not great. It correctly identifies some of the components, but many of the labels are off or pointing in random places. Probably not usable in an educational or professional setting for now, as it’s mostly hallucinating.

Conclusion

This was an interesting exercise to understand better what GPT-4o image generation is capable and incapable of doing. What stood out for me is its ability to add elements to an existing image, like in the second test case. It also appeared clear that the model is struggling with understanding the logic of the images.

From a content creation perspective, the possibility to quickly create images that follow my brand guidelines looks promising. Although the occasional hallucinations and not being to able expect a “correct” result can make the whole process slower.

If we consider how much ChatGPT improved from its release, I think it’s just fair to expect image generation to improve as well, and that we’ll be using these tools more and more in our everyday lives. Many sectors have just been changed forever. Resilience and adaptability might very well be the most valuable skills to navigate this era.

1. Industry news

A new smartest AI on Earth, every month

With the DeepSeek craze cooling down, February was a busy month for all major companies working on AI products. I like the following chart because, besides highlighting each model’s capabilities, it also reflects how quickly different organizations can innovate. According to this chart, xAI seems to have the steepest improvement curve, if the self-reported results of the reasoning version were to hold true.

Let’s have a look at all the major announcements:

• OpenAI released GPT 4.5, temporarily available only to Pro users, while they work on adding more GPUs. I haven’t had the chance to test it yet, but the general sentiment seems to be that the model’s capabilities are underwhelming compared to the massive increase in costs (15-30x higher than GPT-4o). The o-series will probably still be better for most STEM tasks, but 4.5 could apparently thrive in research or creative applications, although it feels hard to justify the price tag. Another interesting update is that as of February 25, Deep Research is available to all users on all paid plans of ChatGPT. I haven’t tested it yet, but I hope to find some time to compare it with Gemini’s, Grok, and Perplexity, and potentially write an article about it.

• Anthropic finally released Claude Sonnet 3.7, together with Claude Code, a command-line agentic coding tool. Sonnet 3.7 is the first “hybrid” reasoning model, in which it will avoid reasoning for very simple queries. When integrating with the API, there’s also the option to have granular control over how much (or how little, or not at all) the AI should think before responding. This is one of the features that I’m most excited about testing, as it allows embedding a reasoning model in an application without making users wait forever for a response.

• xAI released Grok 3, trained on a massive cluster of 200,000 Nvidia H100 GPUs. Elon Musk called it the “smartest AI on Earth”. The claim is mainly based on self-reported results, so we’ll have to wait for independent evaluations to confirm that. Grok 3 will also have its own DeepSearch mode: the first Grok agent able to search the internet and distill information.

  

• Perplexity announced their own Deep Research tool and that they’re working on an “agentic” browser called Comet.

• Google released Co-scientist, a multi-agent AI system designed to support researchers in various aspects of the research process. It is currently only available to selected research organizations.

• Figure, a humanoid robot startup, announced it would break out of their partnership with OpenAI after achieving a breakthrough on their side. They also claimed that they will show “in the next 30 days something no one has ever seen on a humanoid”. In another X post, they unveiled grocery-sorting robots, and announced plans to ship 100k robots within the next four years. I’m excited to never have to do laundry again, but I also have to admit that their videos remind me of one of those movies where it doesn’t end well for humans…

What I’m thinking

Increased competition is bringing us increasingly powerful tools every month. I’ve been working to implement LLM intelligence within Ento for the last few months. I love that as I work to make our application interface better, the models are also improving in parallel. We’re at a level of modularity that makes it incredibly easy for us to switch between different models as smarter, cheaper, and faster versions appear.

At the same time, I feel like we’re lacking an easy and structured way to benchmark different models on our concrete data and tasks. There are plenty of general benchmarks out there, but how does Gemini or o3 fare compared to Sonnet 3.7 for my concrete task? Right now I have to come up with creative and time-consuming ways to compare the outputs, while I would love to have a tool to perform structured comparisons in my own application. If you know of one, let me know!

The 1000 TWh problem

With higher AI capabilities, increasingly powerful models being deployed, and more companies joining the frontier model race, many are concerned about the climate impact of AI. The following figures on projected data center consumption for the US and China can help us get a better understanding of the situation. We’re talking about a potential combined consumption of almost 1000 TWh by 2030, with data centers potentially consuming up to 12% of total US electricity by 2028.

Given these staggering numbers, there’s increased interest in assessing and improving the computational efficiency of LLMs. Hugging Face, in collaboration with Salesforce, announced the AI Energy Score project, a public benchmark aimed at standardizing the measurement of inference compute requirements for different models.

Google also released a detailed report on the carbon emissions of their AI hardware. It’s interesting to see that operational emissions are still responsible for 75-90% of total emissions, even when accounting for the data centers’ construction phase and hardware manufacturing.

Google is still pursuing its ambitious goal to run all of its Cloud on 100% CFE, which is impressive given the developments of the past 2 years. It means that they have a lot of confidence in their ability to improve both hardware and software efficiency while also deploying renewables and getting into favorable PPAs.

What I’m thinking

Data center consumption is definitely something that should be taken seriously. At the same time, I believe we have tremendous room for technological improvement. In the figure below, we can see how the biggest drivers of US electricity consumption are actually buildings and data centers.

For both of them, we can significantly reduce emissions simply by deploying more renewables coupled with (AI-powered) demand-side flexibility solutions. While doing research for this article, I just came across these two publications about “carbon-aware machine learning inference” for image recognition and LLMs. You should give them a read as they’re super interesting.

The AI Energy Score project is also a step in the right direction. Standardizing efficiency benchmarks is very necessary. At the same time, I feel like the quality of the output should be taken into account as well. It’s great to see a ranking of which model is more efficient at inference in terms of GPU compute, but when evaluating which LLM I’m going to embed in my application, I need to be able to see the efficiency together with a measure of the output quality. Much of the model efficiency discussion focuses on the number of tokens processed per unit of energy consumed, but I find this misleading. o1 and GPT-3 could receive the same amount of tokens and output the same number tokens, yet their result can be incredibly different. For me, the ultimate benchmark should be amount of output intelligence (which will vary depending on the type of task) per Wh consumed. We’ll then need to look at the Pareto frontier for the quality vs. cost trade-off.

Lights out in Chile

Many of you will have heard about the blackout that plunged Chile into darkness on February 25th. The lives of many people were disrupted, and there have been at least three confirmed deaths among people relying on electricity to live. But what I want to focus on here is how this situation represents a scenario that will become increasingly common globally, as we keep installing more renewables where the generation potential is highest, while not investing enough in transmission grids. Here are a few interesting facts to consider:

• Chile is leading South America’s renewable energy expansion, with solar PV and wind now comprising around 40% of the country’s installed capacity.

• Chile's installed capacity is set to increase by 139% by 2060, driven predominantly by solar, battery storage, and onshore wind.

• By 2060, over 90% of Chile’s power generation will come from renewables, with solar accounting for nearly half of the total installed capacity.1

• The transmission system infrastructure of Chile is quite limited, with the north (where most PV power plants are) and the center (where most of the demand is) of the country being connected by a single main 500 kV line.

Most solar power production plants are currently concentrated in the north of the country.

Given the photovoltaic power potential of the territory, this is a trend that is likely to continue over the coming years:

Chile’s geography exacerbates the situation, but having high renewable capacity far away from main demand centres with limited transmission infrastructure is a recipe for blackouts. The solution? As usual: invest more in grids, deploy energy storage, and facilitate demand-side flexibility mechanisms.

What I’m thinking

Aside from the technical energy aspects, researching about this blackout got me thinking about how our society has grown extremely dependent on electricity. Many people were stuck in elevators for hours; it took two hours to evacuate the Santiago metro, which transports 2.3m passengers per day; and patients’ care in hospitals was disrupted. This is a reminder that, as much as we care about lowering emissions, ensuring energy security must be the first priority of any energy policy.

Chile is a very fascinating country under many aspects. If you’re interested in learning more, I recommend you read this article from one of my favorite Substack authors. I really enjoyed it because it’s filled with maps, and I love a good map!

2. Scientific publications

1. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers reveals that while AI tools can reduce the perceived effort of critical thinking tasks, they can also lead to over-reliance on AI, diminishing independent problem-solving abilities. The study also shows that higher confidence in AI correlates with less engagement in critical thinking. Another study, Generative AI can harm learning, conducted in a high school math setting, demonstrates that while AI-based tutors can boost immediate performance, removing AI access can lead to worse learning outcomes than if AI was never introduced, suggesting a "crutch" effect.

   

   What I’m thinking

   While I’m an enthusiastic early adopter of many of these technologies, I also think that one of the biggest risks is using them to replace critical thinking altogether. The negative effects of this are clearly visible in our social media feeds, which are getting flooded by bland content that sounds the same. I now typically stop reading as soon at the first “delve”, “pivotal”, or “transformative”.

   Personally, I avoid using AI to write as much as possible. At the same time, AI is boosting my productivity significantly on the coding side. This article aligns well with my thinking about AI-assisted coding, which you can also find in this post.

   It’s evident that more work must be done on designing AI tools that actively promote critical thinking. We need an approach that leverages AI's strengths without compromising fundamental cognitive skills. Mira Murati, the former OpenAI CTO, agrees. Her new startup, Thinking Machines Lab, is reportedly focusing on improving the interaction between humans and AI.

2. SolNet: Open-source deep learning models for photovoltaic power forecasting across the globe introduces an open-source deep learning model generator for PV power forecasting. The architecture uses transfer learning by first pre-training on PVGIS data and then fine-tuning on observational data. The results show RMSE improvements of 10-15% in data-scarce environments.

   What I’m thinking

   In the last Currents issue, we discussed the potential of using transfer learning when training models for automated building control. This month, I was happy to review yet another application where transfer learning is showing great promise. I also really like that this publication is accompanied by reproducible open source code (accessible here). In my opinion, when a scientific paper includes reproducible code, its impact is so much higher.

3. Reimagine Energy publications

In my latest code tutorial I look at how a single energy conservation measure can lead to a 28% reduction of rooftop solar ROI:

Reimagine Energy

More efficient buildings lead to lower rooftop solar returns

Built environment decarbonization is a long and complex process that will keep us busy for much of this century. When organizations first start looking into decarbonizing their portfolios, they quickly realize there’s a limited budget and many potential projects to optimize energy consumption and reduce emissions. A natural question quickly comes up…

Read more

a year ago · 6 likes · Benedetto Grillone

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Internship: Imbalance & Imbalance Price Prediction in the Belgian Electricity Market at ML6

2. Data Scientist - AI Benchmark Architect at PassiveLogic

3. Lead Data Engineer at Octopus Energy UK

4. Energy/Environmental Technology Researcher at Berkeley Lab

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

In this post, I want to exemplify this by showing how the economics of a rooftop PV system can change drastically when an Energy Conservation Measure (ECM) is implemented in a building.

At Ento, I’ve been working for quite some time on a tool to predict the optimal decarbonization roadmap for a certain building or portfolio. One of the challenges with these types of tools is that the order in which projects are implemented is not interchangeable. For instance, consider a building where we could:

1. Convert existing lighting to more efficient LED fixtures

2. Replace a gas boiler used for space heating with a heat pump

3. Upgrade to better-insulated windows

4. Install a rooftop PV system

Each action has a certain impact on final energy consumption and carbon emissions, but they also influence one another:

1. LED installation will reduce the electricity demand

2. Replacing the gas boiler with a heat pump will eliminate gas consumption but increase electricity consumption.

3. Replacing windows with more insulated ones lowers the energy required for space heating.

4. A rooftop PV system produces electricity that can be self-consumed or exported to the grid depending on demand during production hours.

Self-consuming PV-produced electricity is almost always more economically profitable than injecting it into the grid, which means the optimal size for a rooftop system depends heavily on the building’s load profile. Because various measures, such as lighting retrofits or heating system changes, affect that load, we need to carefully consider their impact when designing the rooftop PV system.

To illustrate this, I will consider the same building we’ve used in past tutorials, a large facility located in Washington, DC. In this building, a large ECM was implemented, and through several tutorials, we’ve been evaluating its impact on the building load. It’s also the building we used as an example for calculating the optimal PV size based on the pre-ECM electricity consumption.

In this tutorial, we will consider the same building and simulated production data. The difference is that we will match that data with the post-ECM consumption to see how the electricity self-consumption, self-sufficiency, and the annual return on investment change.

We’ll require a few things to get our final result, but I’ll skip the code snippets that were explained in previous posts. You can find all our previously published tutorials at this link, but if you need a quick recap:

• In tutorials [1], [2], and [3] we built a counterfactual model to estimate energy efficiency savings from an implemented ECM

• In tutorial [4] we extrapolated those savings to one full year from only six months of data

• In tutorial [5] we simulated the solar energy produced by a rooftop PV system installed on the building.

• In tutorial [6] we matched the production with the building consumption to calculate self-sufficiency and self-consumption ratios.

• In tutorial [7] we compared the energy production simulated with pvlib with the one provided from the Google Solar API for the same location

• In tutorial [8] we calculate the optimal capacity of the rooftop PV system based on the pre-ECM consumption data of the building.

One issue with running PV calculations after the ECM implementation is that for this site we don’t have a full year of post-ECM data. To solve this, we had to predict and extrapolate the post-ECM consumption to span an entire year, which we did in tutorial [4]. Here’s the result we got:

If you want to replicate the results from the following lines of code, you’ll need these 3 variables which were defined in previous tutorials:

• module_energy: the hourly production of a single panel, as calculated in tutorial [5]

• predicted_consumption_post_ecm: a full year of post-ECM electricity consumption estimated with a counterfactual model in tutorial [4]

• results_df and kneedle to plot the PV system optimization results, as calculated in tutorial [8]

# we'll use the solar energy production estimated for year 2016, to avoid biasing the calculation because of different weather data
# we need to change the index of the production timeseries from 2016 to 2018, excluding Feb 29th (2016 was a leap year)
module_energy_2018 = module_energy[~((module_energy.index.month == 2) & (module_energy.index.day == 29))]

module_energy_2018.index = module_energy_2018.index.map(lambda x: x.replace(year=2018))

We can now run the PV size optimization again using the post-ECM data.

# we match the PV production to the consumption and calculate the self-consumption ratio and the self-sufficiency ratio

post_ecm_results_df = pd.DataFrame()
for system_size in system_sizes:
    panel_count = system_size / module_rated_power
    print('system_size', system_size, 'panel_count', panel_count)
    # match the PV production to the consumption
    pv_production = module_energy_2018 * panel_count / 1000
    # calculate the grid consumption as the difference between the consumption and the production (but not less than 0)
    grid_consumption = (predicted_consumption_post_ecm - pv_production).clip(lower=0)
    print('grid_consumption', grid_consumption.sum())
    # calculate self consumption (electricity that is consumed on site)
    self_consumption = predicted_consumption_post_ecm - grid_consumption
    # calculate grid injection
    grid_injection = pv_production - self_consumption
    # calculate the return from self consumption
    self_consumption_return = self_consumption.sum() * electricity_price
    # calculate the return from injection
    injection_return = grid_injection.sum() * injection_price
    # calculate the total annual return
    total_return = self_consumption_return + injection_return
    # calculate the self-consumption ratio
    self_consumption_ratio = self_consumption.sum() / pv_production.sum()
    # calculate the self-sufficiency ratio
    self_sufficiency_ratio = self_consumption.sum() / predicted_consumption_post_ecm.sum()
    # add results as a new row to the results dataframe
    post_ecm_results_df = pd.concat([post_ecm_results_df, pd.DataFrame({'system_size': [system_size], 'self_consumption_ratio': [self_consumption_ratio], 'self_sufficiency_ratio': [self_sufficiency_ratio], 'total_return': [total_return]})], ignore_index=True)

We can now plot the PV size optimization results run on the pre-ECM data (from tutorial [8]) and on the post-ECM data.

from plotly.subplots import make_subplots

fig = make_subplots(rows=2, cols=1, subplot_titles=('With ECM', 'Without ECM'), specs=[[{"secondary_y": True}], [{"secondary_y": True}]])

# First subplot (with ECM)
fig.add_trace(go.Scatter(x=post_ecm_results_df['system_size'], y=post_ecm_results_df['self_consumption_ratio'], 
                        mode='lines', name='Self Consumption Ratio (ECM)', line=dict(color='#ffc107')), row=1, col=1, secondary_y=False)
fig.add_trace(go.Scatter(x=post_ecm_results_df['system_size'], y=post_ecm_results_df['self_sufficiency_ratio'], 
                        mode='lines', name='Self Sufficiency Ratio (ECM)', line=dict(color='#ff5733')), row=1, col=1, secondary_y=False)
fig.add_vline(x=kneedle_post_ecm.knee, line_dash="dash", line_color="#7ac53c", annotation_text=f"Optimal Size ({round(kneedle_post_ecm.knee)} kWp)", row=1, col=1)
fig.add_trace(go.Scatter(x=post_ecm_results_df['system_size'], y=post_ecm_results_df['total_return'], 
                        mode='lines', name='Total Return (ECM)', line=dict(color='#007BFF')), row=1, col=1, secondary_y=True)

# Second subplot (without ECM)
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_consumption_ratio'], 
                        mode='lines', name='Self Consumption Ratio', line=dict(color='#ffc107')), row=2, col=1, secondary_y=False)
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_sufficiency_ratio'], 
                        mode='lines', name='Self Sufficiency Ratio', line=dict(color='#ff5733')), row=2, col=1, secondary_y=False)
fig.add_vline(x=kneedle.knee, line_dash="dash", line_color="#7ac53c", annotation_text=f"Optimal Size ({round(kneedle.knee)} kWp)", row=2, col=1)
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['total_return'], 
                        mode='lines', name='Total Return', line=dict(color='#007BFF')), row=2, col=1, secondary_y=True)

# Update layout and axes
fig.update_layout(height=800, showlegend=False)

fig.update_yaxes(title_text="Self Consumption Ratio", secondary_y=False, row=1, col=1)
fig.update_yaxes(title_text="Total Return", secondary_y=True, row=1, col=1)
fig.update_yaxes(title_text="Self Consumption Ratio", secondary_y=False, row=2, col=1)
fig.update_yaxes(title_text="Total Return", secondary_y=True, row=2, col=1)

fig.show()

The profiles of the three curves change significantly before and after the ECM. We can notice a few things by analyzing the plot:

1. The optimal system size decreased, going from 1700 kWp to 1517 kWp

2. Before the ECM, the optimal system yielded a $97k annual return with 69% self-consumption and 27% self-sufficiency. After the ECM, it yields $60k, 40% self-consumption, and 29% self-sufficiency. This means that there was a 28% reduction in return and a 42% reduction in self-consumption for the optimal system.

3. The self-consumption ratio and annual return curves shifted significantly. For instance, if we stuck with a 1700 kWp system post-ECM, we’d only get $63k in return and 36% self-consumption.

While it’s straightforward that a lower overall consumption would lead to a smaller optimal system, the way in which the two curves shift is more nuanced. The shape of the annual $ return curve depends on the difference between the price of self-consumed energy and exported energy. If both prices were equal, the annual ROI line would be a simple linear function of system size. But when self-consumption is valued differently from injection into the grid, the relationship depends heavily on how well the production aligns with the building’s consumption. In the post-ECM case, the shape of the self-consumption ratio curve changed substantially because the ECM had a bigger impact during the hours that coincided with solar production. As we scale up the PV system, more energy ends up being exported to the grid rather than self-consumed.

We can plot the reduction in ROI as a function of the size of the system to get an even better idea of the dynamics:

# plot the % reduction in annual ROI for each system size
return_reduction = (post_ecm_results_df['total_return'] - results_df['total_return']) / results_df['total_return']

fig = go.Figure()
fig.add_trace(go.Scatter(x=post_ecm_results_df['system_size'], y=return_reduction, mode='lines', name='ROI Reduction', line=dict(color='#007BFF', width=10)))
fig.add_trace(go.Scatter(x=post_ecm_results_df['system_size'], y=self_consumption_reduction, mode='lines', name='Self-consumption Reduction', line=dict(color='#ffc107', width=10)))
fig.update_layout(
    xaxis_title="System Size [kWp]"
)
fig.show()

At smaller capacities, both the pre- and post-ECM scenarios are offsetting a large share of on-site consumption at the higher retail rate, so the difference in returns is minimal. As we move up to a medium-sized system, the pre-ECM building can still put more solar output toward meeting its relatively higher midday load, whereas the post-ECM building begins to export more electricity at a lower tariff—widening the gap in returns. Then, for very large systems, both scenarios end up exporting most of their solar production, so the return difference narrows again.

This quick analysis shows how a single ECM can radically change financial projections. If we installed a 1700 kWp system before the ECM implementation expecting an annual return of $97k, but ended up with only $63k instead, that would cause serious problems for many portfolio managers.

Furthermore, this highlights how difficult it is to plan an optimal decarbonization strategy for a real estate portfolio when every project you implement can reshape the ROI of subsequent projects. Solving this problem efficiently requires a powerful blend of computational tools and engineering creativity, and it’s a really interesting challenge to be working on. If you’re working on similar topics, I’d love to chat about different approaches—feel free to reply to this email or connect with me on LinkedIn!

Thanks for reading Reimagine Energy! Subscribe for free to receive new posts and support my work.

1. Industry news

DeepSeek and LLM democratization

This month’s headlines have all been about DeepSeek, the latest LLM from China that caused panic on Wall Street and generated widespread discussion in the tech community. Here’s brief summary of what DeepSeek is:

DeepSeek is an open-source large language model developed by a Chinese research lab that is backed by a major quant/hedge fund. It comes in two flavors:

• V3: A pretrained model that nearly matches the performance of some U.S. frontier models at a fraction of the cost.

• R1: A “reasoning” model that, via reinforcement learning and techniques like Group Relative Policy Optimization, shows chain-of-thought capabilities comparable to early OpenAI models (like o1).

  Thanks for reading Reimagine Energy! Subscribe for free to receive new posts and support my work.

The DeepSeek team employed advanced efficiency engineering (e.g., optimized key–value cache compression, mixture-of-experts architectures)1 to squeeze more performance out of less expensive, lower-grade chips (e.g. Nvidia H800s vs the more recent H100s employed by leading AI companies), demonstrating that high-quality LLMs can now be trained with dramatically lower compute and monetary costs.

So why is DeepSeek such a big deal?

1. Geopolitics: DeepSeek is the first non-U.S. model to come close to state-of-the-art U.S. systems on key benchmarks at a much lower training cost. This challenges the long-held assumption that massive (and expensive) infrastructure is needed to build competitive LLMs. It also raises concerns about the impact of U.S. export controls on advanced chips.

2. AI ecosystem impact: It comes with open weights, released under a permissive MIT license, which has already spurred a flood of derivatives on platforms like Hugging Face2 and may end up leveling the playing field.

3. Business implications: Nvidia stock fell 17% this week amid doubts about whether large-scale investments in compute are actually required. The competitive moat of companies like OpenAI and Anthropic was also questioned.3

4. Energy efficiency: DeepSeek’s training cost and energy use are lower, although widespread deployment and increased accessibility could trigger a rebound effect where greater efficiency drives a surge in overall energy consumption.

   

What I’m thinking

These are the most interesting consequences from the DeepSeek release in my opinion:

1. More competition in the field means better and cheaper options for the end-user. A couple of weeks ago we had to pay a subscription to work with reasoning models. Today we get OpenAI o3-mini, DeepSeek, Microsoft Copilot, and the Google AI Studio Flash Thinking models, all for free.

2. Lower infrastructure and training costs push us in the direction of having commoditized models. The rest of us can then focus on building apps that provide value on top of such models (the so-called “wrappers”). When everyone has access to equally powerful models, the real competitive advantage lies in deeply understanding the unique problems that we’re trying to solve with AI.

3. If I had to pick one technical innovation from the DeepSeek paper, it’s that the model learned to reason on its own, using a variant of reinforcement learning called Group Relative Policy Optimization. Companies building chat interfaces for LLMs have widely applied Reinforcement Learning from Human Feedback to tune the the model’s behavior by showing it good examples of user interaction. DeepSeek seems to have taken the humans out of the loop.

4. The obvious upside of a more efficient model is its reduced cost and energy use, but this deserves its own analysis. I’ll cover that in the next paragraph.

Nadella, Jevons, and more efficient AI

Last Sunday, probably in an attempt to shield Microsoft from the upcoming Wall Street bloodbath, Satya Nadella tweeted about how more efficient models will lead to increased AI adoption. To make his point, he cited the Jevons paradox.

What is the Jevons paradox?

William Stanley Jevons was an English economist who, among other things, studied the impact of Britain’s coal supplies on the British economy. In his 1865 book “The Coal Question”,4 he writes:

Jevons observed that fuel efficiency improvements in steam engines (and later, in other energy technologies) did not lead to lower coal consumption. Instead, as the fuel became cheaper per unit of work, its usage increased—sometimes even exceeding the original consumption levels.

The concept has been later expanded to: “Increased efficiency lowers the cost per unit of energy service, which in turn can stimulate demand and ultimately lead to higher total energy use.”

Does it make sense then to explain increased AI use with Jevons’ ideas?

More efficient AI training and inference leading to increased AI adoption doesn’t seem related to the Jevons paradox. It’s actually what you’d expect from a technology that gets cheaper and better—it will be used more.

Still, the Jevons paradox could be applied to the potential increased electricity use that would be generated from more efficient (hence cheaper and more accessible) AI use. This is what different scientists in the field have been discussing and trying to understand—this paper provides a good overview.

But does this imply negative consequences for humanity?

What I’m thinking

Let’s take a step back. Most of us want a world with lower emissions, but why is that? We observed that emitting uncontrolled amounts of greenhouse gases into the atmosphere could lead to catastrophic environmental consequences. This could potentially lead to an unrecoverable collapse of civilization or even the complete extinction of humanity.

Indeed, the extinction of humanity would be an easy way to reach “Net Zero”. But is that what we want? No. Arguably the final goal of those concerned with decarbonization is for human society to flourish in a sustainable and resilient way.

Electrification, mechanization of agriculture, rail transport, etc. all lead to an increase in greenhouse gas emissions in the past. Would we go back and give up on these advancements? Probably not.

I believe we’re in a similar place now with AI.

Will smarter and more efficient AI models increase global electricity consumption? Possibly. Does this mean humanity is better off by not developing AI or not making it more efficient and accessible? I don’t think so. The challenge is to develop a prosperous energy-rich civilization while also avoiding climate catastrophe.

Trump’s first weeks in office

Donald Trump was inaugurated as the 47th U.S. president on January 20th, with a display of his big tech minions seated in the first row.5 Here’s an overview of Trump’s policy initiatives on Energy and AI during his first 15 days:

1. Declared a national energy emergency to ease environmental restrictions, enabling rapid expansion of fossil fuel projects.

2. Officially withdrew from the Paris Agreement, taking distance from global climate commitments.

3. Rolled back clean energy initiatives, including EV tax credits, offshore wind leasing, and clean hydrogen funding.

4. Issued orders to expand fossil fuel production by resuming offshore and Arctic drilling, fast-tracking LNG export permits, and replenishing oil stockpiles.

5. Announced major projects (notably the “Stargate Project”) with hundreds of billions of dollars allocated to build state-of-the-art AI data centers and infrastructure in the U.S.

6. Launched public-private partnerships with industry giants such as OpenAI, SoftBank, Oracle, and Microsoft to secure U.S. leadership in AI, aiming to create hundreds of thousands of jobs and enhance national security.

   

What I’m thinking

The rapid growth of AI adoption is projected to double or triple data centers’ electricity demand by 20286. Under Trump’s energy policies, the most immediately scalable solution to increase electricity supply is natural gas. Gas power plants are quick to deploy and can provide the reliable baseload power required by data centers. While natural gas burns cleaner than coal, it still emits substantial amounts of carbon dioxide and faces other challenges like methane leakage.

Relying heavily on natural gas will drive up emissions and set back the U.S. in the global race for clean energy. A focus on fossil fuels might, in the short term, “Make America Great Again”. In the long term, this is likely to undermine competitiveness in the emerging green economy, erode U.S. influence in global climate negotiations, and delay the global transition to a sustainable, low-carbon energy system.

2. Scientific publications

1. Transfer learning in building dynamics prediction and A scalable approach for real-world implementation of deep reinforcement learning controllers in buildings based on online transfer learning: The HiLo case study

   These two papers look at how transfer learning (TL) can be used to boost the performance of AI-based HVAC control systems. The first study compares six TL strategies for deep neural network models that predict building dynamics—showing that selectively fine-tuning the decoder layers improves predictive accuracy. The second paper demonstrates a real-world online transfer learning approach for deep reinforcement learning (DRL) controllers. Pre-trained in simulation via digital twins and then transferred between offices, the method combines imitation learning and fine-tuning to outperform traditional rule-based controllers and DRL models trained from scratch.

   What I’m thinking

   I’ve followed advances in transfer learning applied to buildings ever since I started working in this field because they address one of the key issues in data-driven solutions: the lack of data. With more than 30,000 buildings already on the Ento platform, we’re excited about the potential to build a “hive mind” where data from different buildings can be leveraged to continually optimize operations and reduce energy waste globally.

   

2. Agent Laboratory: Using LLM Agents as Research Assistants presents a fully automated, multi-agent research pipeline where specialized LLM agents collaborate to different tasks from literature review and experimental design to coding and report generation, thereby cutting research costs dramatically. In a complementary fashion, Hallucination Mitigation Using Agentic AI Natural Language-Based Frameworks tackles one of the most persistent challenges in generative AI by orchestrating multiple agents that iteratively review, fact-check, and refine content, with novel metrics to quantitatively track the reduction in hallucinations.

   What I’m thinking

   Everyone’s talking about AI agents as a key trend for 2025. Amid the inevitable noise, these two papers show two tangible applications: using agents to accelerate research and reduce hallucinations. Despite the impressive progress, I believe that having a human expert with critical thinking in the loop is still the key to ensure the outputs are reliable and truly meaningful. More on that in the next section.

3. Reimagine Energy publications

I was away most of January, so I didn’t get a chance to write a Reimagine Energy post. But I did have time to read, so here’s some of the content I loved and I think is worth sharing:

1. Nat Bullard’s annual presentation on the state of decarbonization is a must-read for everyone interested in the topic. Some of my favourite slides are 3, 77, 89, 95, and 171.

2. Humanity’s Last Exam: this benchmark on which LLMs still perform poorly, feels more representative of a reality in which I still can’t offload the majority of my daily tasks to AI.

3. I’ve been thinking a lot lately on the impact of AI on writing and human critical thinking7, and I enjoyed reading these Substack posts on the topic:
   

   internet princess

   Choosing to walk

   i get a lot of messages asking me for writing advice, which i’m often hesitant or unable to give because i don’t think i’m particularly good at the actual task of writing: i don’t have a writing schedule, a daily output, good desk posture (or a desk at all), a spiritual artistic practice, or any of that other stuff that people tend to want to hear about…

   Read more

   a year ago · 1217 likes · 5 comments · rayne fisher-quann
   Cosmos Institute

   Essay Writing as Personal Sovereignty

   Over 115 readers entered our inaugural Cosmos Institute Essay Contest, where we asked for reflections on the theme of human autonomy in the age of AI. We received essays from across the globe, with submissions from the UK, Italy, France, Germany, and India. The response was remarkable, and we were impressed by the range of ideas…

   Read more

   2 years ago · 186 likes · 17 comments · Cosmos Institute and Michael Dean

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Data Engineer at Hypercube

2. PhD position in physics-informed data-driven techniques for supporting the energy transition of European cities at Cimne/UPC

3. Artificial Intelligence Engineer at MeteoSim

4. AI Energy Engineer at EKORE Digital Twin Solution

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

1. Industry news

Is infinite energy all we need to reach AGI?

The last weeks of 2024 have been hectic. Let’s look at the major announcements in the space over the past month:

• OpenAI announced a $200-per-month ChatGPT Pro subscription which gives access to o1 pro mode, a version of o1 that uses more compute to think harder and provide even better answers to the hardest problems.

• Text-to-video generators: OpenAI’s Sora was released for ChatGPT Plus subscribers in some regions. Google announced the release of rival product Veo2.

• Google also released Gemini 2.0 Flash Experimental. I’ve been playing with this latest version, and the outputs were, most of the time, on par with o1. Since Gemini is free, it might be the best option if you don’t want to pay for a subscription right now.

• Gemini 2.0 will also power Project Mariner, an agent that can take control of your browser to search for information, fill out forms, and even shop.

• Microsoft released Phi-4, a 14-billion-parameter small language model (SLM), that specializes in complex reasoning and math.

• Amazon announced Amazon Nova, a new family of foundation models that can handle text, images, and videos. They claim that the Nova family will be about 75% cheaper than competing models.

But what caught everyone’s attention was the announcement of OpenAI o3 and the impressive results the model achieved on the ARC-AGI benchmark.

The ARC-AGI is a benchmark designed to assess an AI’s ability to show adaptive general intelligence by solving tasks that are easy for humans but tough for computers. It involves visual puzzles and tasks that require understanding basic concepts like objects, boundaries, and spatial relationships.

OpenAI’s o3 model scored 87.5% on this benchmark, surpassing previous records and achieving performance comparable to human-level intelligence (85% average). For reference, GPT-4o could only achieve 5% on this benchmark at the beginning of 2024. This has been celebrated as a fundamental leap toward Artificial General Intelligence (AGI).

How did o3 achieve such a major improvement? o3 leverages Chain of Thought (CoT) prompting, which encourages models to articulate their thinking by solving problems step by step. o3 then solves tasks generating thousands of Chain of Thoughts (CoTs) and, at test time, searching across them to find the steps needed for the final answer.

This also results in very long compute times and high cost for the model. According to OpenAI, solving each of the ARC-AGI tasks using o3 cost around $3,400 per task, which covers computing, searching, and evaluating possible answers at inference time. This led to total spending of $1.6 million for the entire benchmark.

What I’m thinking

Humans can solve these puzzles in minutes (how long did it take you to solve the one above?). There is a considerable difference between actual intelligence and a system that’s essentially trying every possible reasoning path until it stumbles on the right one.

What we really want is a system that naturally adapts, reasons abstractly, and does so as effortlessly as humans do. Current large language models (LLMs) show some remarkable capabilities: they can appear to reason, solve specialized tasks, and significantly boost productivity in coding and content creation. But they’re still missing crucial human-like qualities such as genuine understanding of logic, or the ability to create knowledge and ideas that were not already embedded in their training set.

Is, then, infinite energy all we need to achieve AGI? LLMs keep improving, and their parent companies are pushing them to their limit to remain relevant and justify their astronomical valuation and cloud costs. Still, my opinion right now is that it’s very unlikely that we’ll reach general intelligence by employing the same architectures and just throwing in more compute time.

What does o3 mean for our energy system?

Even though AGI likely won’t happen this way, companies like OpenAI will keep scaling compute at the inference level. With o3, aside from pre-training compute costs, the cost for each task completion will rise. This will push data center energy demand even higher.

How will we power all of this? We’ll likely need a mix of renewables, batteries, natural gas, and nuclear. Above all, we’ll need to invest heavily in expanding and modernizing our electricity grids.

In the image below, we can see one of the forecasts for the generation capacity in Europe from the Montel EU Energy Outlook 2024, including the total gross demand and inflexible demand.1

Let’s look more closely at the different elements:

Electricity Grid expansion

In the UK, £77 billion were recently committed to expand the electricity grid. This will gradually happen everywhere, as most existing grids are obsolete and undersized when considering the growing electricity demand.2 Analysts say we’ll need to invest about $3.1 trillion for grid expansion and modernization.3

Renewables

After years of surprising (and largely unpredictable) solar installation growth, the EU Market Outlook for Solar Power4 showed a considerable slowdown in the EU in 2024. This might be the new normal for a while—unless we see some major breakthroughs in storage solutions that allow to increase the capture rate.

Germany experienced another Dunkelflaute in December, with power prices shooting up to over 800 €/MWh. The interconnection with the Nordics caused Denmark, Sweden and Norway to also experience these extremely high prices. Germany’s Energiewende is being blamed for recent spikes, due to its nuclear phase-out in favor of high renewable capacity.5

Batteries

Energy storage prices have dropped significantly over the past few years, causing a boom in battery deployments. In the U.S., the grid added battery capacity equivalent to 20 nuclear reactors in just the past four years.6 Meanwhile, in the EU, installed energy storage capacity tripled in 2024 compared to 2023.7

Nuclear

Big tech companies ramped up nuclear investments through the end of 2024.8 Signing deals for power from refurbished plants can make economic sense, but building new ones from scratch takes a long time, and the economics are still unclear. Small modular reactors (SMRs) have been the cool new kid on the block for a while, but they haven’t yet proved they can operate at a competitive price, so they’re unlikely to be a short-term solution.

What I’m thinking

Planning the energy system of the future is a complex task that involves making decisions under many uncertainties. Energy policy planning for governments and utilities is even more complicated because these issues are getting very political: right-wing parties often lean toward keeping fossil fuels to avoid harming the economy, and left-wing parties push for more renewables and clean tech. My current analysis on the energy supply mix for the coming years is this: renewables will get cheaper and cheaper, so we’ll install more wind and solar as it makes the most economic sense. Their intermittency can be problematic, but batteries and equipment that respond to price signals should fix most of these problems. In the long term, baseload will mainly come from nuclear. In the short term, we’ll likely keep using natural gas. This is a complex and fascinating topic. I’m working on a comparative analysis of different energy sources, so stay tuned for that article coming in early 2025.

More AI-driven deals in the energy sector

Trane Technologies will acquire Brainbox AI in a first-of-its-kind deal of this scale in the sector. We already covered Brainbox AI in previous issues, when we discussed the release of its autonomous assistant, ARIA. Brainbox will join Trane keeping its team intact, but leveraging Trane’s partnerships and existing digital building management technology to grow adoption.

In Europe, more funding rounds for companies working on demand-side flexibility were announced:

• Belgian Powernaut raised €2.4 million pre-seed to accelerate Virtual Power Plant adoption.

• Swiss Hive Power raised €3.5 million seed round to scale its AI-based software for optimising electric vehicle charging.

• Belgo-American NOX Energy raised $1 million to integrate smart devices into the energy flexibility market.

What I’m thinking

Brainbox AI’s acquisition is a new milestone in the sector. Although incumbents like Schneider Electric have been acquiring smaller climate-tech startups for years, Brainbox stands out with over 190 employees and a Series A and B that totaled $60 million. On one hand, this deal confirms a strong market demand for AI-powered HVAC control; on the other, it illustrates that even the strongest players in the market might still rely on larger companies for long-term survival.

Meanwhile, recent EU funding rounds reveal that with Europe’s volatile electricity prices, flexibility is the new hot topic for VCs. It’ll be interesting to watch how these companies develop their products. The market exists, and regulation is slowly getting there. Ultimately, success will depend on effective product design, which is the hardest and most fascinating aspect!

2. Scientific publications

1. An interpretable data analytics-based energy benchmarking process for supporting retrofit decisions in large residential building stocks. The authors analysed a dataset of 49,000 Energy Performance Certificates (EPCs) from the Piedmont region in Italy. They used it to estimate and benchmark primary energy demand for space heating and domestic hot water. Different machine learning techniques were used to extract insights from the data. An additional Explainable AI (XAI) layer explained the results to non-experts.

   What I’m thinking

   EPCs are publicly available datasets that haven’t been widely analysed with machine learning. At Ento, we’re also working with thousands of these EPCs for our Danish clients. Having access to both the EPCs and the real consumption data from the buildings, we can identify and try to explain the differences between the theoretical and actual energy performance. The approach from the paper is thorough and worth investigating further. One issue we’ve found, when working with EPCs, is the inconsistent data structure among different regions and countries, and the lack of easy access through an API.

2. How long is long enough? Finite-horizon approximation of energy storage scheduling problems. In this math-heavy paper, the authors examine how most real-world energy storage system schedulers use a finite rolling horizon to manage what is fundamentally an infinite-horizon problem. They derive a lower bound on how long the horizon must be, based on the technical characteristics of the storage system. They also give an upper bound on how “bad” the solution can get if the horizon is too short.

   What I’m thinking

   As mentioned earlier, storage systems are becoming vital pieces in our energy infrastructure, and optimally scheduling them is a complex challenge. These highly mathematical problems are perfect candidates for tools like o1 Pro. In 2025, we’ll likely see more of these solutions in deep scientific research. To be sure: this doesn’t mean we won’t need researchers and domain experts anymore. They’re the only ones who can understand an LLM’s output and verify whether it’s valid. As LLMs reach PhD-level capabilities, research will accelerate, and the level required to do meaningful work will likely increase.

3. Reimagine Energy publications

Check out my latest tutorial to detect the optimal size of a rooftop solar PV system using Python.

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Data Scientist at Aurora Energy Research

2. Chief Technology Officer at myenergi

3. Data Scientist at NOX Energy

4. Marie Curie Industrial PhDs at UPC

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

7

GEM Energy Analytics

The rise of large-scale batteries in Europe

Aside from the rise of solar energy, another significant trend currently unfolding is the proliferation of large-scale batteries worldwide, and in Europe in particular. This trend is poised to have a substantial impact on our grids and the electricity markets…

Read more

a year ago · 47 likes · 9 comments · Julien Jomaux

This tutorial is part of a series on simulating and analyzing rooftop solar energy production using Python. In previous tutorials, we simulated energy production for a roof using pvlib, analyzed electricity flows for a real building from the Building Data Genome Project, and compared physics-based methods for solar estimation to Google’s Solar API.

For brevity, we won’t recalculate the solar production of a single module in this tutorial. You can find those calculations in earlier tutorials, so for the code that follows, we’ll assume the variable module_energy is already available in the workspace.

Let’s start by importing the libraries we need for the analysis.

import pandas as pd
import numpy as np
import plotly.graph_objects as go

Then let’s import the consumption data from the building we’re analyzing. This building is part of the Building Data Genome Project 2. For more details about the dataset and the data extraction process, you can refer to this tutorial.

# read electricity consumption file

meters_df = pd.read_csv('data/electricity_cleaned.csv')

# set the timestamp column as index of the dataframe

meters_df.set_index('timestamp', inplace=True)
meters_df.index = pd.to_datetime(meters_df.index)

# create a new dataframe with only the data the selected building

building_df = meters_df['Rat_education_Alfonso']
building_df = building_df.rename('consumption')

# select one year of consumption data

annual_site_consumption = building_df['2016-01-01':'2016-12-31']

Now that the consumption data is available, we’ll calculate the hourly solar production of a rooftop PV system for multiple system peak capacities. Then, for each system size, we’ll determine the ratio of self-consumption and self-sufficiency.

The maximum system capacity is limited by the rooftop area, which we previously found to be about 22,500 m². Based on the panel’s characteristics (also described previously), this area corresponds to 12,680 panels, or a peak capacity of 5,072 kWp.

# define parameters of the system

module_rated_power = 0.4 # kWp
max_capacity = 5072 # kWp
min_capacity = 0.1 * max_capacity
electricity_price = 0.10 # $/kWh
injection_price = 0.02 # $/kWh

# we want to test 200 different configurations in between the min and max size

num_systems = 200
system_sizes = np.linspace(min_capacity, max_capacity, num_systems)

Let’s now iterate over 200 different configurations and calculate self-sufficiency and self-consumption ratios for each. We define the self-consumption ratio (SCR) as the ratio between the total annual self-consumed energy (Esc) and the total annual energy produced (Ep) by the PV system. This indicates how much of the PV production is used onsite versus how much is injected into the grid.

We define the self-sufficiency ratio (SSR) as the ratio between total annual self-consumed energy (Esc) and total annual building consumption (Ec). It represents how much of the building’s load is covered by solar.

# calculate scr and ssr for each of the 200 system sizes
results_df = pd.DataFrame()

for system_size in system_sizes:

    panel_count = system_size / module_rated_power
    pv_production = module_energy * panel_count / 1000
    
    # calculate the grid consumption (electricity imported from the grid)
    grid_consumption = (annual_site_consumption - pv_production).clip(lower=0)
    
    # calculate self consumption (electricity that is consumed on site)
    self_consumption = annual_site_consumption - grid_consumption

    # calculate grid injection (electricity injected back into the grid)
    grid_injection = pv_production - self_consumption

    # calculate the self-consumption ratio
    self_consumption_ratio = self_consumption.sum() / pv_production.sum()
    
    # calculate the self-sufficiency ratio
    self_sufficiency_ratio = self_consumption.sum() / annual_site_consumption.sum()
    
    # add results as a new row to the results dataframe
    results_df = pd.concat([results_df, pd.DataFrame({'system_size': [system_size], 'self_consumption_ratio': [self_consumption_ratio], 'self_sufficiency_ratio': [self_sufficiency_ratio]})], ignore_index=True)


# plot the results

fig = go.Figure()
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_consumption_ratio'], mode='lines', name='Self Consumption Ratio'line=dict(color='#ffc107')))
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_sufficiency_ratio'], mode='lines', name='Self Sufficiency Ratio', line=dict(color='#ff5733')))
fig.show()

Output

Let’s analyse these results.

• For small PV sizes (left side of the x-axis), nearly all the generated solar energy is used onsite, so self-consumption is close to 100%.

• As PV capacity increases, the building can’t consume all the extra electricity the system generates, so the self-consumption ratio drops. More energy is exported to the grid.

• The self-sufficiency ratio (the fraction of the building’s load covered by PV) is relatively low for smaller systems, because there’s not enough solar capacity to meet the demand.

• As system size increases, PV production covers a larger portion of the total energy needs. However, it eventually plateaus. After about 2,000 kWp, each additional kWp only adds a small incremental increase in coverage.

These metrics can help us understand the system’s dynamics: as we increase capacity, we cover more of the building’s load but self-consume less of the produced electricity. Note that this graph will be different for each individual building analyzed, as it represents how well production hours align with consumption hours for that building.

Although it’s informative, this plot doesn’t give a clear idea of how to determine the optimal system size. Let’s include another factor: payback time, which is often the most important factor when installing a PV system. We’ll calculate the system’s economic benefit by assuming an electricity cost of $0.10/kWh and a compensation of $0.02/kWh for electricity injected into the grid.

# define prices
electricity_price = 0.10 # $/kWh
injection_price = 0.02 # $/kWh

results_df = pd.DataFrame()

for system_size in system_sizes:
    panel_count = system_size / module_rated_power
    
    # match the PV production to the consumption
    pv_production = module_energy * panel_count / 1000
    

    # calculate the grid consumption (electricity imported from the grid)
    grid_consumption = (annual_site_consumption - pv_production).clip(lower=0)
    
    # calculate self consumption (electricity that is consumed on site)
    self_consumption = annual_site_consumption - grid_consumption

    # calculate grid injection (electricity injected back into the grid)
    grid_injection = pv_production - self_consumption

    # calculate the self-consumption ratio
    self_consumption_ratio = self_consumption.sum() / pv_production.sum()
    
    # calculate the self-sufficiency ratio
    self_sufficiency_ratio = self_consumption.sum() / annual_site_consumption.sum()

    # calculate the return from self consumption
    self_consumption_return = self_consumption.sum() * electricity_price

    # calculate the return from injection
    injection_return = grid_injection.sum() * injection_price

    # calculate the total annual return
    total_return = self_consumption_return + injection_return

    # add results as a new row to the results dataframe
    results_df = pd.concat([results_df, pd.DataFrame({'system_size': [system_size], 'self_consumption_ratio': [self_consumption_ratio], 'self_sufficiency_ratio': [self_sufficiency_ratio], 'total_return': [total_return]})], ignore_index=True)

# plot the results

fig = go.Figure()
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_consumption_ratio'], mode='lines', name='Self Consumption Ratio', line=dict(color='#ffc107')))
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_sufficiency_ratio'], mode='lines', name='Self Sufficiency Ratio', line=dict(color='#ff5733')))

# plot the total return on a different y axis
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['total_return'], mode='lines', name='Total Return', yaxis='y2', line=dict(color='#007BFF')))
fig.update_layout(yaxis2=dict(overlaying='y', side='right'))
fig.show()

Output

The growth of total annual economic return also slows down as the system capacity increases. This happens because as more capacity is added, the SCR decreases, and a larger share of electricity is sold back to the grid at $0.02/kWh instead of being self-consumed at $0.10/kWh.

In practical use-cases, a CapEx analysis is typically performed to find the optimal system size, considering the capital investment, annual returns, and how electricity prices and inflation might evolve. We won’t go too deep into the financial details here, but the metrics we calculated can help us build an intuition about the capacity “sweet spot.” This intuition can then support the CapEx analysis. By examining the chart, we see that between 1,000 kW and 2,000 kW, the SSR plateaus and the annual return begins to slow down. Can we find a way to pinpoint this inflection point?

Mathematically, finding this inflection point is often called detecting the “knee” or “elbow” of the function, where the rate of increase starts to flatten. To find this point, we’ll use the Kneedle algorithm, a systematic method for detecting it in monotonically increasing curves. The algorithm was presented in this paper and we’ll use the kneed package for its Python implementation.

# find the knee algorithmically

from kneed import KneeLocator

# apply the Kneedle algorithm to the annual return function

kneedle = KneeLocator(results_df['system_size'], results_df['total_return'], S=1.0, curve="concave", direction="increasing")

print('Optimal system size:', round(kneedle.knee, 0), 'kWp')

Output

Let’s also plot the results:

# plot scr and ssr

fig = go.Figure()
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_consumption_ratio'], mode='lines', name='Self Consumption Ratio', line=dict(color='#ffc107')))
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['self_sufficiency_ratio'], mode='lines', name='Self Sufficiency Ratio', line=dict(color='#ff5733')))

# add a vertical line at the elbow
fig.add_vline(x=kneedle.knee, line_dash="dash", line_color="#7ac53c", annotation_text="Optimal Size")

# plot the total return on a different y axis
fig.add_trace(go.Scatter(x=results_df['system_size'], y=results_df['total_return'], mode='lines', name='Total Return', yaxis='y2', line=dict(color='#007BFF')))


fig.update_layout(yaxis2=dict(overlaying='y', side='right'))
fig.show()

Output

The detected optimal capacity is 1,700 kW. Of course, this is just a rule-of-thumb indication of the inflection area. The results of the Kneedle algorithm also depend on the sensitivity parameter (S), which affects how soon the inflection point is found, so the detected inflection point shouldn’t be treated a strict reference.

Conclusion

In this tutorial, we explored a method to estimate the optimal PV system size by considering metrics like the self-consumption ratio, self-sufficiency ratio, and the annual economic return. Although this is a good first analysis, real-world cost-effectiveness studies are more involved, especially when we consider hourly spot market prices. In the last year, we’ve seen high electricity price volatility and declining solar capture rates. This adds complexity because electricity market prices often depend on solar generation as well.

In a future tutorial, I’ll analyze how variable electricity and falling solar capture rates prices affect the cost-benefit analysis of PV systems.

1. Industry news

Trump is back: what does this mean for Energy and AI?

Donald Trump’s re-election was arguably November’s most globally significant news. But what does this mean for the energy and AI landscape? Trump recently vowed to “dramatically increase baseload power” to slash electricity costs, prevent brownouts, and “WIN the battle for AI superiority.” How will this play out?

Key figures in his administration suggest that the focus will be on fossil fuels over renewables. Chris Wright, an oil and gas CEO who claims “there’s no climate crisis,” has been nominated to lead the Department of Energy. North Dakota Governor Doug Burgum is set to head the National Energy Council, which will be granted sweeping authority over federal agencies to establish American “energy dominance” around the world.

Yet, the landscape is more complex now than during Trump’s first term. ExxonMobil CEO Darren Woods stated that the U.S. should not pull out of the Paris Agreement, claiming this could hinder Exxon’s investments in carbon capture, hydrogen fuel, and EV battery technology—initiatives that rely on federal tax credits established under the Inflation Reduction Act (IRA). Repealing the IRA altogether won’t be straightforward for the Trump administration. Many Republican districts stand to gain from its investments, making a full rollback politically tricky.

What I’m thinking: The last thing we need is to politicize decarbonization efforts. Climate change is a global challenge and requires collective effort, not ideological battles. Instead of getting bogged down in rhetoric and polarized discourse, we should improve policies based on what’s effective. Now more than ever, we need to focus on solutions that make immediate economic sense—those are the ones the market will adopt regardless of political winds.

Trump’s re-election is likely to have a significant impact on U.S. and global energy policy, but it won’t stop economically sensible solutions. As the smart folks at WattCarbon observed: “We may witness a shift from top-down to bottom-up decarbonization, driven by grassroots movements rather than federal mandates. […] With limited funds, maximizing the impact of every dollar spent on decarbonization becomes crucial.”

Europe’s industrial struggles and AI-driven decarbonization

While the U.S. is grappling with policy shifts, Europe faces its own energy challenges. Swedish battery maker Northvolt has filed for Chapter 11 bankruptcy, entering a restructuring phase to stay afloat. This is a blow to Europe’s ambitions in energy storage—a critical component for electrifying transportation and balancing renewable energy supply and demand.

Germany is also navigating a difficult situation. Early November brought a “Dunkelflaute”, a period of minimal wind and sunlight, leading to a spike in electricity prices and increased reliance on fossil fuels. While the grid remained stable, electricity costs and carbon footprints rose sharply. This added to existing doubts about the nation’s Energiewende efforts and comes amid political instability. German Chancellor Olaf Scholz’s coalition government collapsed amid disputes over economic recovery and the energy transition. With the finance minister sacked and snap elections set for February, crucial policies now hang in the balance—including the 2025 budget, without which many climate and industry support programs could stall.

Amid these challenges, the German startup scene has been notably active, with three companies raising capital to improve energy efficiency through AI:

• Predium, a Munich-based PropTech company, raised €13 million in Series A funding. They use AI, satellite imagery, and 3D modeling to help property owners assess ESG risks and prioritize renovations for maximum cost savings and subsidies.

• Juna.ai, a Berlin startup, secured $7.5 million to develop AI systems that make industrial manufacturing self-learning and adaptive. Their technology optimizes processes in real-time, improving efficiency and reducing energy consumption and emissions.

• KUGU, another German PropTech firm, raised €8.5 million in Series A funding. They offer products that optimize and control energy systems, helping real estate companies meet emissions targets. KUGU manages over 12,000 buildings and serves major housing companies, utilities, and metering service providers across Germany.

What I’m thinking: Despite the hurdles of a continent struggling from an industrial and deep-tech manufacturing standpoint, Europe’s software push for bottom-up decarbonization offers hope. Multiple software startups leveraging AI to improve energy efficiency and reduce operating costs demonstrate that innovation doesn’t wait for political stability. Similar to the U.S., the way to navigate these moments of policy uncertainty is to focus on economically viable products, rather than betting solely on technologies that depend on rebates and subsidies.

Other projects and reports that caught my attention

I enjoyed DeepMind’s latest report “A new golden age of discovery”, which provides an overview of how AI is transforming scientific disciplines. The report emphasizes AI’s potential to tackle many challenges in the energy sector as well, such as creating simulations to advance fusion energy research or improve weather forecasting. I also appreciated the discussion around creativity. LLMs are trained to minimize anomalies, and I’ve noticed this in my own experience. I rely on LLMs as my “sparring partners” when debating ideas and exploring different points of view, but I rarely use them during the idea generation phase—the output is usually bland. We definitely shouldn’t rely solely on LLMs to define the areas toward which we conduct our research, at least for now.

I also looked at FlexiDAO’s report on granular meter data accessibility. I talk a lot about the importance of granular energy data and how, without it, AI-powered solutions can’t optimize our energy systems effectively. The report is comprehensive, and the different map visualizations provide a clear way to understand which markets founders in this space should focus on and which emerging markets are on the rise.

Finally, I explored Robin Hawkes’ GB Renewables Map. I found this through a LinkedIn post about how a brand-new 1.1 GW wind farm in the UK has been “off” nearly 70% of the time during 2024 due to limitations on the transmission system between Scotland and England. The analysis led me to Robin’s GB Renewables Map, and it was really fun to explore—I suggest you check it out!

What I’m thinking: I love that every month brings exciting new products, reports, whitepapers, and interactive platforms that open new doors and unlock new connections. There’s so much to explore, with new developments showing up every day. On the other hand, the abundance of possibilities and shiny new tools can be distracting and even generate anxiety. As professionals in this sector, we need to master the ability to discern the trivial many from the vital few. Understanding that not all efforts are equal and that some activities yield exponentially higher results than others is crucial for prioritizing effectively.

2. Scientific publications

1. Large Language Models for the Creation and Use of Semantic Ontologies in Buildings: Requirements and Challenges. This paper studies how Large Language Models (LLMs) can aid in constructing and querying semantic models in the building domain, specifically using the Brick schema. Semantic ontologies like Brick offer a standardized, machine-readable framework for representing complex building systems, but creating and querying these models is challenging and often requires specialized expertise. The paper explores how LLMs can address these challenges by lowering the expertise barrier and making the creation and use of semantic models more accessible.

   What I’m thinking: I feel we haven’t seen enough LLM applications in the buildings industry compared to other fields, so I’m glad that research like this is starting to emerge. At the same time, I would have liked to see a more practical application of these ideas. The paper is quite theoretical, and it would have been helpful to see how this approach could work in the reality of building operations and management. Hopefully the next step!

2. Boosted Conformal Prediction Intervals. This study introduces a boosted conformal procedure designed to improve conformal prediction intervals by targeting specific properties like enhanced conditional coverage or reduced interval length. Through gradient boosting, the authors systematically refine a predefined conformity score function without altering the underlying trained model. This approach operates post-training, relying solely on model predictions. Experiments demonstrate that starting from conventional conformal methods, the boosted procedure achieves substantial improvements in reducing interval length and decreasing deviation from target conditional coverage.

   What I’m thinking: This paper is math-heavy and isn’t directly focused on the energy sector, but I found it particularly interesting. Gradient boosting is a powerful tool, and the ability to provide accurate confidence intervals for predictions is valuable across all industries, including ours. In the context of Measurement and Verification of energy efficiency savings, for example, being able to calculate precise confidence intervals for estimated savings enhances the reliability of our results. This method could potentially improve how we assess and communicate the uncertainty in our energy savings estimates, leading to better decision-making and increased trust from stakeholders.

3. Reimagine Energy publications

If you didn’t already, make sure to check out my Python tutorial to predict a full-year of energy efficiency savings when only 6 months of data are available:

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Data Scientist at WattTime

2. Data Scientist and Machine Learning Engineer - Energy and Food at Eurecat

3. Senior Data Scientist at Danfoss

4. Principal Research Scientist - AI for Power Systems at Siemens Energy

5. Postdoctoral researcher: Combining AI, adaptive control, and digital twins to optimize autonomous systems at Zurich University of Applied Sciences

Conclusion

With so much going on in the sector, it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.), please reply to this email or reach out on LinkedIn and I’ll be happy to consider them for inclusion!

This happened to me recently. I wanted to analyze how solar panel savings change when combined with other energy conservation measures. So I set out to combine a full year of on-site panels’ production with the building’s consumption before and after implementing an ECM. The issue was that I only had 6 months of post-implementation data and I needed a full year to make accurate comparisons.

So I asked myself: Could we predict the next 6 months of consumption based on the historical data from the building and on the patterns we see in the first 6 months of post-ECM data?

In this tutorial, I'll show you how to use LightGBM to extrapolate your partial-year energy savings into a full-year estimate. We'll build a model that learns from your building's energy patterns and predicts future consumption.

Getting Started

First, we need to import the necessary libraries:

import pandas as pd
import lightgbm
import plotly.graph_objects as go

Loading electricity consumption data

We’re using data from the Building Data Genome Project 2. For details on this dataset and how to handle it, have a look at our previous tutorial:

# Read electricity consumption file

meters_df = pd.read_csv('data/electricity_cleaned.csv')

# Set the timestamp column as index of the dataframe

meters_df.set_index('timestamp', inplace=True)
meters_df.index = pd.to_datetime(meters_df.index)

# Create a new dataframe with only the data for the selected building

building_df = meters_df['Rat_education_Alfonso']
building_df = building_df.rename('consumption')

Loading and preprocessing weather data

We’ll need weather data for 2016, 2017, and 2018. I downloaded this data from the National Solar Radiation Database, but any historical weather data source will do.

# Process and combine 3 years of weather data files (2016-2018)
# Each file needs special handling since first row contains metadata:
# 1. Drop first row which contains metadata
# 2. Use the new first row as column headers
# 3. Reset index to avoid duplicate indices when concatenating

# The function below handles this preprocessing for each weather file
def process_weather_file(filepath):
    df = pd.read_csv(filepath)
    df = df.iloc[1:].copy()  # Drop first row and make copy
    df.columns = df.iloc[0]  # Set column names from first row
    return df.iloc[1:].reset_index(drop=True)  # Drop the header row and reset index

# Process all weather files
weather_files = [
    'data/washington_dc_weather_2016.csv',
    'data/washington_dc_weather_2017.csv', 
    'data/washington_dc_weather_2018.csv'
]
weather_dfs = [process_weather_file(f) for f in weather_files]

# Concatenate all dataframes
weather_df = pd.concat(weather_dfs)

# Create datetime index
weather_df['datetime'] = pd.to_datetime(weather_df[['Year', 'Month', 'Day', 'Hour', 'Minute']])
weather_df.set_index('datetime', inplace=True)

# convert values to float
weather_df[['Temperature', 'Wind Speed', 'Wind Direction', 'Pressure', 'Dew Point']] = \
    weather_df[['Temperature', 'Wind Speed', 'Wind Direction', 'Pressure', 'Dew Point']].astype(float)

# select only relevant columns
weather_df = weather_df[['Temperature', 'Wind Speed', 'Wind Direction', 'Pressure', 'Dew Point']]

# resample to hourly
weather_df = weather_df.resample('H').mean()
weather_df.head()

Output

Merging consumption and weather data

Let’s merge the consumption data with the weather data and add features needed for the analysis.

We know from the previous tutorial that the energy conservation measure was implemented on 2017-06-28, let’s add a boolean feature marking the pre- and post-installation periods.

# merge the dataframes keeping all rows from weather_df by using right merge
df = building_df.to_frame().merge(weather_df, left_index=True, right_index=True, how='right')

# Add time-based features
df['hour'] = df.index.hour
df['dayofweek'] = df.index.dayofweek
df['week'] = df.index.isocalendar().week.astype('int')

# add a new column "energy_conservation_measure" which is 1 after the ECM implementation date
df['energy_conservation_measure'] = 0
df.loc[df.index >= pd.to_datetime('2017-06-28'), 'energy_conservation_measure'] = 1

df.head()

Output

In our previous tutorial on counterfactual energy modeling, we had already built a model to estimate the adjusted baseline consumption after the ECM implementation as if the ECM had not been implemented. You can check it out here:

The post-ECM installation data available for this building spans only half a year (from mid 2017 until the end of the year). Our goal is then to estimate what the consumption will be for the year 2018, if it follows the same trend as after the ECM implementation. We’ll also need to calculate an adjusted baseline consumption for the year 2018 to estimate the savings. Let’s define the datasets we need for our analysis.

Defining analysis periods

We need to define key dates and periods for our analysis:

• Pre ECM Period: Before ECM implementation (before June 21, 2017)

• Installation Period: June 21–28, 2017

• Initial Reporting Period: Post-installation until the end of 2017

• Full Reporting Period: All of 2018

# Define key dates for analysis

installation_start = pd.to_datetime('2017-06-21')
installation_end = pd.to_datetime('2017-06-28')
last_consumption_date = pd.to_datetime('2017-12-31')

training_data_pre_ecm = df[df.index <= installation_start]
training_data_full_period = df[df.index <= last_consumption_date]
prediction_data_post_ecm = df[df.index >= installation_end]
prediction_data_2018 = df[df.index > last_consumption_date]

Baseline model function

We’ll use LightGBM to build our baseline model.

# define function to build baseline model

def build_baseline_model(data, target_variable, features, seed=None):

    data = data.dropna(subset=[target_variable])
    X = data[features]

    y = data[target_variable]
    dataset = lightgbm.Dataset(X, y)

    objective = "l2"
    lgbm_params = dict(
        boosting_type="gbdt",
        n_jobs=1,
        objective=objective,
        seed=seed or 42,
        verbosity=0,
    )
    cv = lightgbm.cv(
        params=lgbm_params,
        train_set=dataset,
        stratified=False,
        num_boost_round=200,
        nfold=5,
        callbacks=[lightgbm.early_stopping(5)],
    )

    # Full model (use best num-rounds from cross-validation)

    model = lightgbm.train(
        params=lgbm_params,
        train_set=dataset,
        num_boost_round=len(cv[f"valid {objective}-mean"]),
    )

    # Predict

    y_pred = pd.Series(model.predict(X), index=X.index)

    return {
        "data": data,
        "features": features,
        "target": target_variable,
        "seed": seed,
        "params": lgbm_params,
        "estimator": model,
        "cv": cv,
        "preds": y_pred,
    }

Predicting consumption for 2018

Now, we’ll predict the building’s consumption for the year 2018.

# Define features and train model

features = df.columns.tolist()
features.remove('consumption')

baseline_model = build_baseline_model(
    data=training_data_full_period,
    target_variable='consumption',
    features=features
)

# Make predictions on reporting data
predicted_consumption_post_ecm = pd.Series(
    baseline_model['estimator'].predict(prediction_data_2018[features]), 
    index=prediction_data_2018.index
)

Visualizing the results

Let’s plot the actual consumption and the predicted consumption for 2018.

# plot the actual consumption and the predicted consumption for 2018

fig = go.Figure()

fig.add_trace(go.Scatter(x=training_data_full_period.resample('1D').sum().index, y=training_data_full_period.resample('1D').sum()['consumption'], line=dict(color='#007BFF')))

fig.add_trace(go.Scatter(x=predicted_consumption_post_ecm.resample('1D').sum().index, y=predicted_consumption_post_ecm.resample('1D').sum(), line=dict(color='#ff5733')))

fig.show()

Output

The model seems to perform well, the 2018 predictions align closely with the consumption patterns observed after the ECM implementation in 2017.

Calculating the adjusted baseline consumption

Next, we’ll calculate the adjusted baseline consumption for the full year 2018, representing what consumption would have been without the ECM.

# Estimate adjusted baseline consumption for the post-installation period

adjusted_baseline_model = build_baseline_model(
    data=training_data_pre_ecm,
    target_variable='consumption',
    features=features
)

adjusted_baseline_consumption_post_ecm = pd.Series(
    adjusted_baseline_model['estimator'].predict(prediction_data_post_ecm[features]), 
    index=prediction_data_post_ecm.index
)

Visualizing all results

Let’s plot the actual consumption, adjusted baseline consumption, and predicted consumption for 2018.

# Plot the actual consumption, the adjusted baseline consumption, and the predicted (post-ECM)consumption for 2018
fig = go.Figure()

fig.add_trace(go.Scatter(x=adjusted_baseline_consumption_post_ecm.resample('1D').sum().index, y=adjusted_baseline_consumption_post_ecm.resample('1D').sum(), line=dict(color='#7ac53c')))

fig.add_trace(go.Scatter(x=training_data_full_period.resample('1D').sum().index, y=training_data_full_period.resample('1D').sum()['consumption'], line=dict(color='#007BFF')))

fig.add_trace(go.Scatter(x=predicted_consumption_post_ecm.resample('1D').sum().index, y=predicted_consumption_post_ecm.resample('1D').sum(), line=dict(color='#ff5733')))

fig.show()

Output

Calculating full-year savings

Finally, we’ll calculate the estimated electricity savings for a full year after the ECM implementation.

# Calculate savings for a full year after the ECM implementation

real_consumption_after_ecm = prediction_data_post_ecm['consumption'].sum()

predicted_consumption_one_year_after_ecm = predicted_consumption_post_ecm[predicted_consumption_post_ecm.index < (installation_end + pd.DateOffset(years=1))].sum()

adjusted_baseline_one_year_post_ecm = adjusted_baseline_consumption_post_ecm[adjusted_baseline_consumption_post_ecm.index < (installation_end + pd.DateOffset(years=1))].sum()

full_year_savings = adjusted_baseline_one_year_post_ecm - predicted_consumption_one_year_after_ecm - real_consumption_after_ecm

print(f"Estimated electricity savings for a full year after the ECM implementation: {full_year_savings:.0f} kWh")

Output

Conclusion

With a few lines of code, we’ve extrapolated six months of energy savings to estimate a full year’s worth. This approach is useful when data is limited or when you want to project future savings soon after implementing an ECM.

Although the methodology is powerful in its simplicity, there are a few key things we also need to take into account:

1. Seasonality: If savings are seasonal and not all seasons are represented in your post-installation data, this method may not be reliable.

2. Non-routine events: The method assumes no significant changes in building consumption patterns (e.g., operational changes, additional ECMs).

3. Future weather data:  If predicting future consumption without actual weather data, you can consider using historical averages or climate models.

For a more detailed analysis on using boolean variables to mark ECM implementation periods, check out this scientific paper that was part of my PhD research.

What’s next?

In my next tutorial, I'll use these full-year estimates to analyze how rooftop solar installation benefits change when combined with other energy conservation measures.

I hope to see you in the next tutorial! If you have any questions, feel free to leave a comment or reach out.

They say good things come to those who wait! This month’s Currents is a week late but filled with exciting developments in AI and Energy.

1. Industry news

The agentic reasoning era has begun, accompanied by a surge in energy demand

The release of OpenAI’s o1 marked a shift in the world of large language models, now moving from “thinking fast” to “thinking slow.” This advancement offers more powerful AI capabilities, but also comes with a considerable increase in energy consumption for each query. What does this mean for our energy landscape?

1. Recent data from the International Energy Agency’s World Energy Outlook 2024 highlight a surge in global electricity demand. Over the next decade, we’re set to add the equivalent of Japan’s annual electricity consumption to the grid each year. “In energy history, we’ve witnessed the Age of Coal and the Age of Oil – and we’re now moving at speed into the Age of Electricity” said Fatih Birol, the IEA’s executive director.

2. While AI holds immense potential for humanity, some experts express concern that “generative AI is making the hardest problem we’ve ever had to solve that much harder to solve,” as noted in this MIT Technology Review article.

3. The need of large AI players to quickly access power generation is also creating social and health concerns. In Southwest Memphis, for example, xAI is building what aims to be the world’s largest supercomputer. But without approval to draw from the main grid yet, they deployed natural gas-powered mobile generators, further contributing to local pollution.

What I’m thinking: I’m a techno-optimist with respect to these issues. AI is propelling humanity at an exponential pace and I believe we can find technological solutions to the energy demand problem. In fact, most of the solutions are already available, we just haven’t found an effective way to deploy them at scale. We’re just at the dawn of the LLM age, recent research suggests we can reduce the energy cost of LLMs by 95%, and more and more research will come out on this topic. While it’s true that AI’s energy demand is creating immediate challenges, we can’t turn back now. Pandora’s box is open. All we can do is roll up our sleeves and get to work.

The supply debate: renewables vs. nuclear

The surge in energy demand has increased the focus on energy supply as well. I’ve gathered insights to understand the current state and future of energy production:

1. Last month, the last British coal power plant closed. The UK’s National Grid System Operator claims that blackout risks are very low, thanks to newly installed wind capacity and interconnections with Denmark. However, challenges remain, as wind lulls can last for days, posing stability issues for the electricity system.

2. An insightful analysis by fellow Substack author Thomas Pueyo explores solar energy’s potential. Projections suggest we’re at the beginning of the S-curve for solar electricity. Its growth is accelerating, and the 2020s through 2030s will see a massive increase in installed capacity. The cost of solar energy is dropping as planned; it may shrink by eight times in the coming decade.

3. However, a key challenge for renewables is their decreasing market value. As more solar capacity is installed, oversupply during sunny periods drives electricity prices down. This makes new solar installations less attractive and existing projects less profitable.

4. Because of their decreasing market value, some energy experts are arguing that wind and solar are the energy past, while the future is nuclear power. While I don’t fully agree with their analysis, I believe it’s important to consider alternative perspectives for informed energy policy decisions in the coming decades.

What I’m thinking: Even after the latest surge in energy demand, I still believe the solution lies in energy efficiency, renewables, storage, and demand-side flexibility. Nuclear will play a role but can’t be the complete solution. We need to cut carbon emissions today, and most nuclear plants take over ten years from commissioning to completion. We’re already too late to rely solely on nuclear to power our future while keeping global temperatures below catastrophic levels. While technological advances in nuclear could happen in the next 5-10 years, I wouldn’t bet humanity’s future on that actually happening.

Unlocking AI’s potential in the energy sector

Every month, I keep an eye out for AI solutions being deployed in the energy sector:

1. In the U.S., utilities are beginning to recognize the potential of AI to support operations and enhance resilience. They’re now looking at predicting outages and extreme weather events. However, accessing data from these utilities remains a challenge. Data that could enable demand-side flexibility programs or reduce consumption through digital products is hard to come by. Utilities often make data access difficult because they aren’t interested in lowering demand; in fact, the opposite may be true. I found a great analysis of this phenomenon in this Substack post by the founder of the Mission:data coalition, Michael Murray.

2. Despite these hurdles, progress is still being made, month after month. Edo and EPRI have earned an award from the California Energy Commission to advance Virtual Power Plants (VPPs) in schools.

3. Last month I found out about WattTime, an exciting project that provides an API to access marginal operating emissions rates—the emissions rate of electricity generators responding to changes in load on the local grid at a certain time. With last month’s data expansion, they now cover 210 countries.

What I’m thinking: If we want to make our energy system smarter, we need to facilitate data access. This is the only way to achieve cost-effective decarbonization. One of the largest barriers for software and AI companies wanting to work with energy is data accessibility. Regulation is lagging behind, but many bottom-up projects are advancing this from various angles. It’s fantastic that the data provided by WattTime is now available at a global scale. While the market might not be ready for this now (most of our customers at Ento are rather interested in an annual average CO2 emissions factor), it soon will be. Tools like this will form the infrastructure upon which we build the future energy system’s applications.

2. Scientific publications

1. Reliably estimating the impact of an active Control strategy in a building. This article introduces a new method to measure energy savings from building control strategies. Traditional measurement and verification methods can take up to two years and are often affected by unrelated changes in building performance. This innovative approach uses a randomized switchback design, frequently alternating between baseline and intervention strategies at fixed intervals—like daily switches. By employing sequential testing and predefined stopping criteria, the method provides reliable energy savings estimates much faster than conventional techniques. In a real-world case study in Chicago, the researchers detected an 11% annual energy savings within 45 weeks, with results remaining consistent even after adding more data.

   What I’m thinking: At Ento, we also struggle to accurately quantify savings from applying our control strategies. Implementing this randomized switchback design might allow us to demonstrate energy savings to our clients much more quickly.

2. Diffusion Model Predictive Control. Google Deepmind researchers introduced Diffusion Model Predictive Control (D-MPC), an approach that combines diffusion models with Model Predictive Control. By learning multi-step action proposals and dynamics models using diffusion techniques, D-MPC outperforms existing model-based planning methods and rivals state-of-the-art reinforcement learning algorithms on standard benchmarks like D4RL.

   What I’m thinking: The ability of D-MPC to quickly adjust to new objectives and changing conditions without retraining aligns perfectly with the challenges we face in optimizing energy usage in real-time. I’m looking forward to seeing how the algorithm performs in real-world buildings.

3. Reimagine Energy publications

Check out my latest article comparing Google’s Solar API with traditional physics-based modeling for solar PV generation:

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Energy Data Science Intern at EnergyHub

2. Data Scientist: Energy System Modeler - PhD at Microsoft

3. Founding Product Engineer at ElectronX

4. Senior Data Engineer at Grundfos

5. PhD position in Control of heating, ventilation, and air conditioning systems of a building at IMT Atlantique

Conclusion

With so much going on in the sector it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.) please reply to this email or reach out to me on LinkedIn and I’ll be happy to consider them for inclusion!

I previously wrote a tutorial on estimating a building’s solar potential using pvlib, a physics-based simulation tool. With Google’s recent release of their Solar API, I couldn’t resist comparing the two methods. Could Google’s API really make solar estimation as easy as an API call?

In this comparison, I set out to evaluate two main factors:

1. Area of the Roof Suitable for Solar Panels: The total surface area where panels can be installed.

2. Production Potential per Panel: How much energy each panel can produce, considering factors like orientation, tilt, and local weather conditions.

Setting the Stage: Selecting a Rooftop in Washington DC

For this comparison, I chose a large rooftop in Washington DC. In the previous tutorials we’ve been analysing the consumption from a large facility located in Washington DC. We don’t have the building’s exact location, so I arbitrarily selected a large site in Washington DC with the help of Google Maps. The selected site is the Walter E. Washington Convention Center.

First Variable: Roof Area Suitable for Solar Panels

Manual Calculation Using Google Maps

By inspecting the site on Maps, I identified two main roof segments suitable for solar panels:

I then used the Google Maps measure distance tool to estimate the total area available for the panels.

# Area available for solar panels according to manual calculations

area_segment_1 = 6000
area_segment_2 = 16500
total_roof_area = area_segment_1 + area_segment_2
print(f'Total area available for solar panels with manual calculation: {total_roof_area} m²')

Output

Automatic Calculation Using Google Solar

Next, I used Google’s Solar API to see what area it detects as suitable.

# Google Solar API call

import requests

api_key = 'your_api_key'
lat = 38.9052
lng = -77.0230
solar_request = requests.get(
    f"https://solar.googleapis.com/v1/buildingInsights:findClosest?location.latitude={lat}&location.longitude={lng}&requiredQuality=HIGH&key={api_key}"
)
solar_info = solar_request.json()

# Area available for solar panels according to the API
print(f'Area available for solar panels according to the Google Solar API: {round(solar_info["solarPotential"]["maxArrayAreaMeters2"])} m²')

# Percentage difference between the two calculations
percentage_difference = ((solar_info['solarPotential']['maxArrayAreaMeters2'] - total_roof_area) / total_roof_area) * 100
print(f'Percentage difference between the two calculations: {round(percentage_difference, 2)}%')

Output

The area detected by the API is almost 30% larger than our manual measurement. To understand better where the difference is from, let’s plot the bounding boxes of the roof segments detected by Google Solar.

# Plot bounding boxes for detected roof segments

import plotly.graph_objects as go

# Prepare data
bounding_boxes = [roof_segment['boundingBox'] for roof_segment in solar_info['solarPotential']['roofSegmentStats']]
data = []
for bbox in bounding_boxes:
    sw = bbox['sw']
    ne = bbox['ne']
    data.append({
        'lat': [sw['latitude'], sw['latitude'], ne['latitude'], ne['latitude'], sw['latitude']],
        'lon': [sw['longitude'], ne['longitude'], ne['longitude'], sw['longitude'], sw['longitude']]
    })
    
# Create the figure
fig = go.Figure()


for d in data:
    fig.add_trace(go.Scattermapbox(
        lat=d['lat'],
        lon=d['lon'],
        mode='lines',
        fill='toself',
        line=dict(width=2, color='blue'),
        fillcolor='rgba(0, 0, 255, 0.3)',
        name='Roof Segment'
    ))

# Set up the map layout
fig.update_layout(
    mapbox=dict(
        style='white-bg',
        layers=[
            dict(
                sourcetype='raster',
                source=[
                    'https://server.arcgisonline.com/ArcGIS/rest/services/World_Imagery/MapServer/tile/{z}/{y}/{x}'
                ],
                below="traces"
            )
        ],
        center=dict(lat=lat, lon=lng),
        zoom=17
    ),
    margin={"r":0,"t":0,"l":0,"b":0}
)

# Display the map
fig.show()

Output

The API’s detected areas don’t align perfectly with the actual roof segments. Some of the identified areas are actually occupied by technical installations, and a large portion of the roof is not being detected.

Second Variable: Number of Panels Fit on the Roof

Google Solar’s estimate

The Solar API provides the maximum number of panels that can be installed.

# Maximum number of panels according to Google Solar
max_panels_api = solar_info['solarPotential']['maxArrayPanelsCount']
panel_capacity = solar_info['solarPotential']['panelCapacityWatts']
panel_height = solar_info['solarPotential']['panelHeightMeters']
panel_width = solar_info['solarPotential']['panelWidthMeters']

print(f"Maximum number of panels (Google Solar API): {max_panels_api}")
print(f"Panel Capacity: {panel_capacity} Wp")
print(f"Panel Dimensions: {panel_height} m x {panel_width} m")

Output

Google estimates we can fit a maximum of 14,648 400 Wp panels on this roof.

Pvlib’s estimate

To make a fair comparison, I selected a similar 400 Wp SunPower panel model. In order to access it, I had to download the latest CEC modules and inverter database, available on the NREL Github.

import pvlib

# Get panel and inverter models

cec_modules = pvlib.pvsystem.retrieve_sam(
    path="data/CEC Modules.csv"
)
cec_inverters = pvlib.pvsystem.retrieve_sam(
    path="data/CEC Inverters.csv"
)

# Choose a 400W panel and inverter
module = cec_modules["SunPower_SPR_MAX3_400"]
module_width = 1.69
module_length = 1.05
inverter = cec_inverters["SunPower__SPR_A400_H_AC__240V_"]

# Calculate how many panels fit in the area identified initially
module_area = module_width * module_length
max_panels_manual = round(total_roof_area/module_area)
print(f"Maximum number of panels (manual estimation): {max_panels_manual}")

Output

Comparing Annual Energy Production

Finally, let’s compare the annual energy production estimated with these two methods.

To run the pvlib calculations, we’ll suppose that the weather_df, temperature_model_parameters, and location variables are already available from the previous tutorial. Let’s simulate the production of one 400 Wp module in Washington DC with pvlib. The Google Solar API assumes that solar panels are installed flush with the roof surface. This means that the inclination of the panels is the same as the pitch of the roof segment they are mounted on. I selected a 5 degrees tilt for the pvlib system, since from the satellite image it doesn’t look like the roof has a significant inclination.

from pvlib.pvsystem import PVSystem
from pvlib.modelchain import ModelChain

# Define PV system characteristics
system = PVSystem(
        surface_tilt=5,
        surface_azimuth=180,
        module_parameters=module,
        inverter_parameters=inverter,
        temperature_model_parameters=temperature_model_parameters,
    )

# Create and run PV Model
mc = ModelChain(system, location, aoi_model="physical")
mc.run_model(weather=weather_df)
module_energy = mc.results.ac.fillna(0)

The Google Solar API returns a list of configurations with different numbers of panels. Let’s compare the expected annual solar production as estimated with pvlib and by the API.

import pandas as pd

# Get the list of configurations proposed by the Solar API
panel_configs = solar_info['solarPotential']['solarPanelConfigs']

# Select configurations evenly
step = len(panel_configs) // 50
selected_configs = panel_configs[::step]
# Add the last configuration if it's not already included
if panel_configs[-1] not in selected_configs:
    selected_configs.append(panel_configs[-1])

# Extract panel counts and yearly energy production for each configuration
panel_counts_selected = [config['panelsCount'] for config in selected_configs]
yearly_energy_selected = [config['yearlyEnergyDcKwh'] for config in selected_configs]
inverter_efficiency = 0.95  # Assume 95% efficiency

# Create a DataFrame
data = {
    'panels_count': panel_counts_selected,
    'solar_api_estimated_production': [energy * inverter_efficiency for energy in yearly_energy_selected],
    'pvlib_estimated_production': [(panel_count * module_energy / 1000).sum() for panel_count in panel_counts_selected]
}
df_configs = pd.DataFrame(data)

# Compare the production estimated with the two methods for different configurations

fig = go.Figure()

fig.add_trace(go.Scatter(
    x=df_configs['panels_count'],
    y=df_configs['solar_api_estimated_production'],
    mode='markers',
    marker=dict(
        size=10,
        color='#ffc107',
        opacity=0.8
    ),
    name='Google Solar'
))

fig.add_trace(go.Scatter(
    x=df_configs['panels_count'],
    y=df_configs['pvlib_estimated_production'],
    mode='markers',
    marker=dict(
        size=10,
        color='#007bff',
        opacity=0.8
    ),
    name='PVLIB'
))

fig.update_layout(
    xaxis_title='Number of Panels',
    yaxis_title='Annual Energy Produced',
    showlegend=True
)

fig.show()

Output

The Google Solar estimation provides a considerably higher estimated production, compared to the physical modelling method. Let’s quantify the difference for the configuration with the maximum number of panels.

# Calculate the percentage difference between Solar API and PVLIB estimates for the configuration with the highest number of panels

max_panels_count = df_configs['panels_count'].max()

# Fetch the solar production estimated by the Google Solar API for the maximum panels count
solar_api_production = df_configs[df_configs['panels_count'] == max_panels_count]['solar_api_estimated_production'].values[0]

# Fetch the solar production estimated by PVLIB for the same maximum panels count
pvlib_production = df_configs[df_configs['panels_count'] == max_panels_count]['pvlib_estimated_production'].values[0]

# Calculate the percentage difference between Google Solar API and PVLIB estimates
percentage_difference = ((solar_api_production - pvlib_production) / pvlib_production) * 100

# Print the results
print(f"Total annual energy production estimated by Google Solar API: {solar_api_production:.2f} kWh")
print(f"Total annual energy production estimated by PVLIB: {pvlib_production:.2f} kWh")
print(f"Percentage difference between Google Solar API and PVLIB estimates at max panel count: {round(percentage_difference, 1)}%")

Output

Although this is a preliminary analysis, it definitely raises some questions about the accuracy and reliability of Google’s Solar API for detailed solar potential assessments. Some potential reasons for the discrepancy might be:

• Weather Data Differences: We used 2016 weather data in the pvlib calculations, while Google Solar might be using a Typical Meteorogical Year or other data.

• Panel Model Variations: While we tried to match the panel specifications, there might be differences affecting efficiency.

• Assumptions in Shading and Losses: Google’s API might be making different assumptions about shading from nearby buildings, system losses, or other factors that influence production.

While the area detection difference might be due to computer vision errors—which is understandable in automated processes—the significant discrepancy in energy produced is concerning. It’s hard to justify a 74% difference based solely on data inputs or minor model variations.

Conclusion

Our comparison reveals significant differences between the two methods:

• Roof Area Detection: The area detected by Google is almost 30% higher than our manual estimation.

• Production Estimates: Google’s estimated annual production for the largest PV system configuration is is 77% higher than pvlib’s.

These findings suggest that while Google’s Solar API offers a convenient way to estimate solar potential, it might not provide accurate results for detailed assessments. Relying solely on the API could lead to overestimations and impact investment decisions.

What’s Next?

Is this discrepancy an isolated case? To find out, in the next tutorial I’ll test more buildings in different locations. I’m also interested in investigating whether Google Solar considers shading in its estimations, and have a look at the financial analyses provided by the API.

Want to help?

Try using the Google Solar API on a building you’re interested in and compare the results with a physics-based model like pvlib. Share your findings with me—I’d be happy to include your analyses in the next issue of the newsletter!

1. Industry news

1. The race for AI supremacy is fueling an unprecedented scramble for energy. Microsoft plans to reopen Pennsylvania’s dormant Three Mile Island nuclear plant by 2028, buying all its power for 20 years. Sam Altman pitched the White House on building massive 5GW data centers—each consuming as much power as three million homes. Y Combinator startup Lumen Orbit suggests moving data centers into space to harness unlimited solar energy. Meanwhile, Meta is researching Reinforcement Learning applications to optimize data center cooling, cutting energy and water usage. There’s also increased interest in edge computing, pushing AI processing closer to where data is generated, reducing latency and easing the strain on central data centers.

   What I’m thinking: OpenAI and Microsoft seem to believe that we haven’t reached the limits of transformer architectures yet, and want to exponentially scale up compute (Mark Zuckerberg seems to agree). The predictions on the energy demands of data centers are so bold that some people believe the economics of operating data centers in space might become viable. Demand response will have a massive role to play here. Edge computing might be part of the solution, with Apple apparently embracing this route as well.

2. San Francisco-based Voltus has unveiled AI Adjuster, a new service that boosts demand response earnings for customers. It uses AI to adjust participation in real-time, tapping into untapped load flexibility. Meanwhile in Europe, Beebop.ai raised $5.5 million seed round to integrate residential devices into the power system. Their software turns household items like storage units, heat pumps, and EV chargers into assets that support the grid.

   What I’m thinking: Over the next decade demand response will become the new normal for all energy users. Companies in both the U.S. and Europe are pushing ahead, even though the market structure is in development. Different countries have different regulations, making things tricky. But those who figure out the tech now will be way ahead once the regulatory framework settles.

   

3. Schneider Electric has unveiled the Building Decarbonization Calculator, an online tool that helps building owners and operators quickly assess energy and carbon conservation measures. Drawing from nearly 500,000 building performance models, the calculator offers recommendations on construction, energy efficiency, and decarbonization strategies. It’s designed to aid compliance with strict regulations like New York City’s Local Law 97, which will impose fines on buildings exceeding emissions limits starting in 2025. The tool also helps users prioritize retrofit plans based on carbon reduction and financial return on investment.

   What I’m thinking: This caught my eye since I’ve been leading development for a similar product, Ento Strategy, over the past year. If a giant like Schneider is diving into this space, it signals a huge market potential. The real game-changer for me will be the user experience. The success of these tools will depend on how they can connect with various data sources and support multiple certifications and regulations.

   

2. Scientific publications

1. Toward global rooftop PV detection with Deep Active Learning. The authors developed a Deep Active Learning method to detect photovoltaic panels in satellite and aerial images. By intelligently selecting which images to label, they reduced the required dataset by 97% compared to traditional methods. The approach uses uncertainty-based algorithms to focus on the most informative images, boosting model accuracy while saving time and resources.

   What I’m thinking: At Ento we developed algorithms to detect PV systems by analysing the electricity time series data of a facility. Being able to cross-reference this with the results of an aerial image analysis could be very interesting and would definitely boost confidence in the results achieved. I’d love to test out this model.

2. Deep learning-based electricity price forecasting: Findings on price predictability and European electricity markets. This article introduces a deep learning-based toolkit designed for day-ahead electricity price forecasting in Europe. It uses a multilayer perceptron model that recalibrates daily and optimizes hyperparameters annually, outperforming traditional models in both speed and accuracy. Findings reveal that while electricity prices have become more volatile due to events like COVID-19 and the global energy crisis, this volatility does not necessarily equate to unpredictability when using advanced deep learning models.

   What I’m thinking: Highly volatile (but easily predictable) power prices are unlocking the deployment of AI-powered demand-side flexibility solutions. Interestingly, time series forecasting remains a domain where transformer architectures haven’t yet become dominant. Instead, traditional models like multilayer perceptrons (MLPs) continue to offer simplicity and computational efficiency while delivering state-of-the-art results.

3. Reimagine Energy publications

The second part of my Python tutorial on integrating building energy consumption and rooftop solar generation data went out last week:

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Lead Data Engineer at Eneco

2. Head of Data Science at Yottar

3. Senior Data Scientist Energy and Utilities at Deloitte

4. Postdoctoral Scholar in Machine Learning and Data-Driven Optimization and Analysis at Future Institute Cities (University of Waterloo)

Conclusion

With so much going on in the sector it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.) please reply to this email or reach out to me on LinkedIn and I’ll be happy to consider them for inclusion!

In this tutorial, we will analyze the interaction between a building's energy consumption and a rooftop photovoltaic (PV) system using Python. This is the second tutorial in our series focusing on rooftop solar systems and their integration with building energy consumption data to determine optimal system sizing and cost-effectiveness. The first tutorial can be found here.

We will continue working with the same facility we analyzed in the counterfactual energy models tutorial, located in Washington, D.C.

Loading the Consumption Data

First, let’s load the building’s electricity consumption data and calculate the monthly totals. The details of the following data loading process are available in the counterfactual energy models tutorial.

import pandas as pd

# Load consumption data

meters_df = pd.read_csv('data/electricity_cleaned.csv')
meters_df.set_index('timestamp', inplace=True)
meters_df.index = pd.to_datetime(meters_df.index)
building_df = meters_df['Rat_education_Alfonso'].rename('consumption')

# Let's have a quick look at the monthly consumption data

monthly_consumption = building_df.resample('M').sum()
monthly_consumption.head()

The monthly consumption for this building is around 260 MWh per month, which is very high. We’re likely looking at a large manufacturing facility or a group of buildings sharing one electricity meter.

Estimating the Required PV System Size

Next, we'll estimate the size of a PV system needed to meet the building's consumption using a simple engineering calculation.

From our last tutorial, we have the monthly production of a PV system in this location, and the system’s peak capacity. We can calculate the average kWh produced per kWp installed capacity for each month of the year. We’re assuming that the monthly production and system capacity variables are available from the previous tutorial.

# Calculate kWh produced per kWp of installed capacity each month
kwh_per_kwp = round(monthly_production / system_peak_capacity)
kwh_per_kwp = kwh_per_kwp.rename('monthly_kwh_per_kwp')
kwh_per_kwp.head()

By comparing kwh_per_kwp with the building’s monthly consumption, we can estimate the required PV system size for each month.

# Calculate required PV system size for each month

monthly_data = pd.concat([monthly_consumption, kwh_per_kwp], axis=1)
monthly_data['required_size'] = round(monthly_data['consumption'] / monthly_data['monthly_kwh_per_kwp'])
monthly_data.head(12)

The table above shows the PV system size needed each month to match the building’s consumption. Since we’re using monthly data, this is a rough estimate with limitations. Even if monthly production matches consumption, they might not align in time.

To begin with, let’s base the system size on the average required capacity across the year. Note that since these calculations are done on monthly data, this is just a rough estimation based on a flawed reasoning. Production and consumption could happen at completely different times. But we need some criteria to decide on a size, and this is a good place to start.

# Specifications for the panel we chose in the previous tutorial 
panel_height = 1.95
panel_width = 0.99
panel_peak_power = 295
tilt_angle = 35
orientation = 180

# Calculate the area occupied by the PV panel on a flat roof
panel_area_flat_roof = (
        panel_height
        * panel_width
        * math.cos(tilt_angle * math.pi / 180)
    )

# Calculate panel count and required roof area
system_capacity_kwp = monthly_data['required_size'].mean()
panel_count = system_capacity_kwp * 1000 / panel_peak_power
roof_area = panel_count * panel_area_flat_roof

print(f"According to the energy consumption matching method, {int(panel_count)} panels are required, covering {roof_area:.1f} m² of roof area.")

This is a very large system, that would typically be found in large-scale commercial or industrial installations.

Analyzing Consumption and Production Interaction

Now, let’s calculate the monthly production of this PV system. We’ll use the module energy data calculated in the previous tutorial, equivalent to the hourly energy output of a panel.

We'll calculate the building's grid consumption and export, after taking into account the PV production.

# Calculate system production, assuming module_energy is available from previous calculations
system_production = panel_count * module_energy / 1000  # Convert Wh to kWh

# Combine consumption and production data
data = pd.concat([building_df, system_production], axis=1).dropna()
data.columns = ['consumption', 'pv_production']

# Calculate grid import and injection
data['grid_flow'] = data['pv_production'] - data['consumption']
data['grid_import'] = data['grid_flow'].clip(upper=0)
data['grid_injection'] = data['grid_flow'].clip(lower=0)

• grid_import represents energy imported from the grid (when consumption exceeds production)

• grid_injection represents excess energy sent back to the grid (when production exceeds consumption)

Results Visualisation

Let’s visualize the daily energy consumption before and after considering PV production.

# Daily data
daily_data = data.resample('D').sum()

# Plot daily consumption with and without PV
fig = go.Figure()
fig.add_trace(go.Scatter(x=daily_data.index, y=daily_data.consumption, line=dict(color='#7ac53c')))
fig.add_trace(go.Scatter(x=daily_data.index, y=-daily_data.grid_import, line=dict(color='#007BFF')))

fig.add_trace(
        go.Scatter(
            x=daily_data.index,
            y=daily_data.consumption,
            mode="lines",
            showlegend=False,
            line=dict(color="rgba(135, 197, 95, 0.2)", width=0),
            fill="tonexty",
            fillcolor="rgba(135, 197, 95, 0.2)",
        )
    )

fig.show()

This line chart compares the daily energy consumed from the grid before and after installing the PV system.

Let’s also look at the monthly energy flows to get a comprehensive view.

# Plot monthly breakdown of energy consumed, produced, and exported

fig_monthly_breakdown = go.Figure()

fig_monthly_breakdown.add_trace(
    go.Bar(
        x=data.resample("M").sum().index,
        y=-data["grid_import"].resample("M").sum(),
        name="Grid Consumption",
        marker_color="#007BFF",
    )
)

fig_monthly_breakdown.add_trace(
    go.Bar(
        x=data.resample("M").sum().index,
        y=data["pv_production"].resample("M").sum(),
        name="Solar Production",
        marker_color="#ffc107",
    )
)

fig_monthly_breakdown.add_trace(
    go.Bar(
        x=data.resample("M").sum().index,
        y=data["grid_injection"].resample("M").sum(),
        name="Excess",
        marker_color="#9185ff",
    )
)

fig_monthly_breakdown.update_layout(
    barmode="group",
    xaxis_title="Date",
    yaxis_title="kWh",
    template="plotly_white",
)

fig_monthly_breakdown.show()

This grouped bar chart provides a clearer picture of how energy is consumed, produced, and exchanged with the grid each month.

Calculating Self-Consumption Metrics

Let’s compute some key metrics for the system.

# Calculate key metrics for the PV system

total_consumption = data['consumption'].sum()
total_production = data['pv_production'].sum()
grid_import = -data['grid_import'].sum()
grid_injection = data['grid_injection'].sum()
self_consumption = total_consumption - grid_import

self_consumption_percentage = (self_consumption / total_production) * 100
coverage_percentage = (self_consumption / total_consumption) * 100

print(f"Self-consumed energy: {self_consumption:.0f} kWh")
print(f"Percentage of total energy production that is self-consumed: {self_consumption_percentage:.1f}%")
print(f"Percentage of total energy consumption that is covered by PV self-consumption: {coverage_percentage:.1f}%")

Even with a large PV system, a substantial amount of electricity is still imported from the grid. This is because PV production doesn’t always match consumption times. Solar energy is generated during the day, but consumption is likely peaking at other times.

Estimating the Payback Time

Finally, let’s estimate how long it would take for such a system to pay for itself. We need to make some assumptions here:

• Electricity price: $0.10 per kWh

• Grid injection compensation: $0.05 per kWh

• System installation cost: $1,000 per kWp

# Financial calculations

electricity_price = 0.10
grid_compensation = 0.05
installation_cost_per_kwp = 1000

total_system_cost = system_capacity_kwp * installation_cost_per_kwp

annual_savings = self_consumption * electricity_price

annual_income = grid_injection * grid_compensation

total_annual_benefit = annual_savings + annual_income

payback_time = total_system_cost / total_annual_benefit

print(f"Total system cost: ${total_system_cost:,.0f}")
print(f"Annual electricity savings: ${annual_savings:,.0f}")
print(f"Annual income from excess energy: ${annual_income:,.0f}")
print(f"Estimated payback time: {payback_time:.1f} years")

Conclusion

In this tutorial, we modeled the interaction between a building’s energy consumption and a rooftop PV system using Python. By estimating the required system size and analyzing performance metrics, we gained valuable insights into the potential benefits and challenges of integrating solar energy into building operations.

Despite installing a large PV system, we observed that a significant amount of energy is still imported from the grid due to the mismatch between production and consumption times. Furthermore, with current assumptions of installation costs, electricity price, and grid injection compensation, the payback time for this system is quite long. Energy storage solutions could enhance self-consumption rates and improve the system's cost-effectiveness.

To answer the initial question: powering this facility entirely with solar energy would require an extremely large system and massive energy storage. This may not be financially viable. Alternative approaches, like shifting consumption patterns or offsetting carbon emissions through Energy Attribute Certificates (EACs), might be better options to reach carbon neutrality.

For this tutorial, we mainly based our calculations on monthly consumption data available from the facility. In the next tutorials, we’ll explore a more complex sizing strategy that considers the building’s specific consumption patterns. This approach will help us identify optimal system capacities that guarantee lower payback times.

1. Industry news

1. Bosch will acquire the residential and light commercial HVAC business from Johnson Controls and Hitachi. This is Bosch’s largest acquisition ever. The revenue from the Home Comfort division of the company is expected to reach €9 billion, about 10% of Bosch Group’s total revenue. Last month Bosch also announced the acquisition of German building automation company DMS AG.

   
   What I’m thinking: Bosch sees the potential in combining HVAC manufacturing with building automation. The sector is changing fast and is ready for disruption. It will be interesting to see how much Bosch can innovate as a large company.

2. Ento connects to the Enedis data hub, gaining access to electricity data from 95% of the country’s meters.

   
   What I’m thinking: Electricity data hubs are extremely powerful. With a single integration, we can access historical and real-time data for millions of meters, enabling AI-powered energy optimization for thousands of businesses and public organizations. Countries without an electricity data hub need to act quickly to avoid missing out on innovation. Allowing building owners and third parties to monitor electricity use might be one of the most cost-effective national investments to decarbonize the building sector.

   

3. Boston startup Cala Systems announces AI-powered heat pump water heater. It is built to maximize energy flexibility by taking into account weather data and learning usage patterns.

   What I’m thinking: Water heating represents 18% of home energy use in the US1, meaning that innovation in this area can make a big impact. These heaters act as thermal energy storage devices, providing flexibility to both households and the grid. Operating a fleet of these systems could also unlock new opportunities for manufacturers.

4. Report from LA startup ChargerHelp shows how poor data is hurting EV charging in the US. Key findings of the report include a critical need for data accessibility and standardization across the industry. This lack of standardization also complicates issue resolution, since software and communication problems cause a large portion of charger failures. You can access the full ChargerHelp report here.

   
   What I’m thinking: The shift to carbon-free mobility is tough. Extracting battery materials, manufacturing vehicles, building charging stations, and decarbonizing the grid are all hardware-intensive tasks with physical limits. We can’t easily change these constraints. But software is different—it can be improved without many of those barriers. That’s why we must minimize software issues and ensure reliable data and applications to support the transition. The report also stresses the need for technicians who understand systems, handle data, and manage QA testing in the field. This aligns with our earlier thoughts: technicians in the energy sector will need to be skilled in working with AI tools and data applications going into the future.

   

2. Scientific publications

1. A probabilistic model for real-time quantification of building energy flexibility. The authors of this paper developed an analytical solution to directly quantify the energy flexibility that can be provided by individual buildings. They then enhanced this with probabilistic techniques and sampling-based uncertainty analysis to determine the flexibility distribution of a cluster of buildings.

   What I’m thinking: Though the study looks promising, it depends on thermal parameters from buildings, that are often hard to access. One solution could be to enhance the model with algorithms that regularly analyze consumption data to estimate and update these parameters. It would be exciting to see this tested in a real-world pilot.

3. Reimagine Energy publications

I was on my annual leave last month, so I didn’t get a chance to write a Reimagine Energy post. But I did have time to read, so here are three Substack posts I loved: a great explanation of negative electricity prices in Europe, a deep dive into using clustering techniques to find patterns in energy consumption data, and a sobering look at the generative AI landscape by music historian Ted Gioia.

GEM Energy Analytics

The reasons for negative prices

In a recent piece where I was cited, The Economist wrote (emphasis added…

Read more

2 years ago · 47 likes · 15 comments · Julien Jomaux

Bricks To Bytes

Decoding Building Energy Patterns

In the built environment, understanding how energy is consumed is crucial for driving efficiency and sustainability. Clustering building energy data helps us identify patterns and trends that reveal where optimization opportunities lie…

Read more

2 years ago · 5 likes · Giuseppe Pinto

The Honest Broker

Google Thinks Beethoven Looks Like Mr. Bean

At first, I thought it was a prank…

Read more

2 years ago · 391 likes · 79 comments · Ted Gioia

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Senior AI Engineer at Enersee

2. International PhD Mines Paris/INESC in AI & Energy

3. AI Consultant at Energy Systems Catapult

4. PhD in Solar Forecasting and Future Sustainable Power Systems at University of Nottingham

Conclusion

With so much going on in the sector it’s not easy to follow everything. If you know of anything relevant that should be included in Currents—such as job posts, scientific articles, or industry events—please reply to this email or reach out to me on LinkedIn. I’ll be happy to consider them for inclusion!

1. Industry news

1. Flexidao launched an hourly renewable energy certification scheme pilot in Japan. Flexidao is a Barcelona-based startup that offers an advanced data, software and advisory solution to the electricity market. They will work with Japanese company Jera Cross to provide their customers with real-time visibility into renewable electricity generated.

   What I’m thinking: Flexidao’s latest pilot highlights a growing global trend towards more granular tracking and certification of renewable energy. Granular tracking of renewable energy generation serves as a critical foundation for developing AI-driven strategies that can ensure clean energy consumption around the clock. The resulting insights from these programs enable the creation of sophisticated demand-side flexibility mechanisms that can optimize energy use in tandem with renewable generation patterns.

   

2. Amazon meets 100% renewable energy goal, while Google is no longer claiming to be carbon neutral. Amazon’s claim is based on an annual matching between energy consumed and clean energy purchased, while the previous neutrality claim from Google was based on carbon offsets.

   What I’m thinking: Annual matching is a good start, but in order to claim real carbon neutrality, hourly matching between consumption and clean production is necessary. Given the increasing focus on AI’s energy impact, and heightened scrutiny of companies' scope 3 emissions, the battle for cloud computing supremacy might be won by the organization that manages to provide certified carbon-free computing power at an hourly level. Technologies like those developed by Flexidao could play a crucial role in helping cloud providers achieve true carbon neutrality.

   

3. Powerledger launches a blockchain-powered energy trading platform in Austria. The platform will allow prosumers to monitor their energy usage and production, and share surplus energy with others across the country.

   What I’m thinking: This project exemplifies how exponential technologies such as blockchain and AI can move beyond buzzwords to deliver real-world value. I’m also a big fan of the community-centric approach. In an increasingly digital world, technology should foster connection rather than isolation. Innovations that bring people together have a higher likelihood of widespread adoption and should be actively encouraged. Ultimately, the most impactful technologies are those that enhance our connections and improve our collective well-being.

   

4. Virtual Power Plants from Tesla and Sunrun delivered hundreds of MW of flexibility to the California grid, as scorching heat put it under extra stress during the month of July. These events involved mainly households with installed solar plus storage systems.

   What I’m thinking: It’s great to see technology helping us adapt to the challenges posed by climate change. As renewable energy and battery storage capacity expands globally, we'll increasingly witness events like this. However, the current approach of simply turning off devices or coordinating multiple batteries to discharge simultaneously is quite basic. The future of VPPs lies in more sophisticated, continuous, and holistic grid management strategies that can adapt in real-time to changing conditions and optimize for multiple objectives simultaneously. Solving the complexity of this system will be one of the most exciting engineering challenges of the next decade.

   

2. Scientific publications

1. A Perspective on Foundation Models for the Electric Power Grid. Many papers have been published over the last couple of years about the use of foundation models for time series. I found this article to be a good overview of their use in the energy sector.

   What I’m thinking: In fields like natural language processing and image generation, foundation models have widely demonstrated their value over the past two years. These areas benefit from well-defined, vast, and accessible datasets, as well as clear metrics for evaluating model performance. Additionally, the risk associated with errors in these applications is generally lower compared to the energy sector, where inaccuracies can have severe consequences. The energy sector’s complexity, regulatory environment, and the critical nature of its operations mean that any new technology, including FMs, must undergo rigorous validation and demonstrate unequivocal reliability before widespread adoption. Nevertheless, they are a powerful tool that we should keep experimenting with. Their application in the power sector represents a fascinating potential solution to an incredibly complex engineering problem.

3. Reimagine Energy publications

Check out my latest code tutorial to simulate rooftop solar energy production using Python.

4. AI in Energy job board

This space is dedicated to job posts in the sector that caught my attention during the last month. I have no affiliation with any of them, I’m just looking to help readers connect with relevant jobs in the market.

1. Senior Data Scientist, Electric Load Forecasting at National Grid

2. Staff Machine Learning Engineer, Autobidder at Tesla

3. Senior Machine Learning Engineer at Carbon Re

4. PhD Candidate in Real Implementation of Model Predictive Control for Building Heating at Norwegian University of Science and Technology

Conclusion

With so much going on in the sector it’s not easy to follow everything. If you’re aware of anything that seems relevant and should be included in Currents (job posts, scientific articles, relevant industry events, etc.) please answer to this email or reach out to me on LinkedIn and I’ll be happy to consider them for inclusion!

We will be working with the same building we analyzed in the counterfactual energy models tutorial. This building is located in Washington DC and has energy consumption data for the years 2016 and 2017.

First, let’s download weather data for Washington DC for the year 2016. There are multiple platforms and APIs that let us access historical weather data. In this case, I used the National Solar Radiation Database from the NREL. Here’s how to do it:

1. Go to the NREL database link.

2. Enter “Washington, District of Columbia” as the location.

3. Select the “USA & Americas (30, 60min / 4km / 1998-2022)” dataset

4. Choose the following variables to include in the dataset: GHI, DNI, DHI, Temperature, Relative Humidity, Wind Speed, and Wind Direction.

5. Remember to mark the “Convert UTC to local time” checkbox.

For those who want to replicate this specific tutorial without downloading the data from the platform, here’s the XLSX file I used.

washington_dc_weather_2016

429KB ∙ XLSX file

Download

Download

Now that we have all the data we need, let’s import the necessary libraries for this tutorial. We will use the latest pvlib version, which is 0.11.0.

import math
import pandas as pd
import plotly.graph_objects as go

import pvlib
from pvlib.modelchain import ModelChain
from pvlib.pvsystem import PVSystem
from pvlib.location import Location

Let’s inspect the weather file we have downloaded.

weather_df = pd.read_excel('data/washington_dc_weather_2016.xlsx')
weather_df.head()

It looks like some data preprocessing is required to remove unnecessary rows and columns. We’ll also want to create a datetime index for our analysis.

#drop the first row and change the header
weather_df = weather_df.drop(0)
weather_df.columns = weather_df.iloc[0]
weather_df = weather_df.drop(1)

# create new datetime column using the columns year, month, day, hour, minute
weather_df['datetime'] = pd.to_datetime(weather_df[['Year', 'Month', 'Day', 'Hour', 'Minute']])
weather_df.set_index('datetime', inplace=True)
weather_df.index = pd.to_datetime(weather_df.index)

# resample to hourly
weather_df = weather_df.resample('H').mean()

# rename columns to use the pvlib nomenclature
weather_df = weather_df.rename(
            {
                "Temperature": "temp_air",
                "Wind Speed": "wind_speed",
                "Relative Humidity": "humidity",
                "Precipitable Water": "precipitable_water",
                "GHI": "ghi",
                "DNI": "dni",
                "DHI": "dhi",
            },
            axis=1,
        )

# convert values to float
weather_df['temp_air'] = weather_df['temp_air'].astype(float)
weather_df['wind_speed'] = weather_df['wind_speed'].astype(float)
weather_df['humidity'] = weather_df['humidity'].astype(float)
weather_df['precipitable_water'] = weather_df['precipitable_water'].astype(float)
weather_df['ghi'] = weather_df['ghi'].astype(float)
weather_df['dni'] = weather_df['dni'].astype(float)
weather_df['dhi'] = weather_df['dhi'].astype(float)

# select only relevant columns
weather_df = weather_df[['temp_air', 'wind_speed', 'humidity', 'precipitable_water', 'ghi', 'dni', 'dhi', ]]

# resample to hourly
weather_df = weather_df.resample('H').mean()
weather_df.head()

The weather dataframe looks much cleaner now. With the weather data ready, we can now define a PV system. We will use the Znshine 295 Wp panel model and an ABB 300W microinverter . While newer models can reach up to 400 Wp, this tutorial focuses on demonstrating the method, allowing the specific panel model and inverter to be easily changed later. For this tutorial, we will assume the system has a 35-degree tilt angle and the panels are oriented south (180 degrees).

# retrieve the inverter and panel specifications from the pvlib library

cec_modules = pvlib.pvsystem.retrieve_sam("cecmod")
sapm_inverters = pvlib.pvsystem.retrieve_sam("cecinverter")
module = cec_modules["Znshine_PV_Tech_ZXP6_72_295_P"]
inverter = sapm_inverters["ABB__MICRO_0_3_I_OUTD_US_208__208V_"]
temperature_model_parameters = pvlib.temperature.TEMPERATURE_MODEL_PARAMETERS[
    "sapm"
]["open_rack_glass_glass"]

# Create a Location and a PV System
location = Location(
    latitude=38.9072,
    longitude=-77.0369,
    name="Washington DC",
    altitude=0,
    tz='US/Eastern',
)

system = PVSystem(
        surface_tilt=35,
        surface_azimuth=180,
        module_parameters=module,
        inverter_parameters=inverter,
        temperature_model_parameters=temperature_model_parameters,
    )

Now we can simulate the electricity generated by a single panel over one year at this location. The only parameter that we need to specify here, apart from the system characteristics and the location, is the angle of incidence (AOI) model. This is the loss model used to account for the reduction in solar irradiance due to the angle at which sunlight strikes the solar panels. In this case we opt for a physical model to calculate AOI losses based on the material properties of the panel’s surface.

# Create and run PV Model
mc = ModelChain(system, location, aoi_model="physical")
mc.run_model(weather=weather_df)
module_energy = mc.results.ac.fillna(0)

# Plot the estimated energy produced by a single panel

fig = go.Figure()
fig.add_trace(go.Scatter(x=module_energy.index, y=module_energy, mode='lines'))
fig.update_layout(yaxis_title='Energy Produced (kWh)')
fig.show()

The pattern suggests seasonal variations, with slightly higher hourly energy production in the warmer months and lower in the colder months, plus some outlier peaks in November.

Let’s now suppose we have an area of 400 m2 that we want to cover with solar panels. Assuming their inclination is 35 degrees, we can calculate how many panels will fit in this area.

# Define panel dimensions and peak power for Znshine_PV_Tech_ZXP6_72_295_P

panel_height = 1.95
panel_width = 0.99
panel_peak_power = 295
tilt_angle = 35
orientation = 180

# Calculate the area occupied by the PV panel on a flat roof
panel_area_flat_roof = (
        panel_height
        * panel_width
        * math.cos(tilt_angle * math.pi / 180)
    )

# calculate amount of panels that fit in a certain roof area
roof_area = 400
panel_count = math.floor(roof_area / panel_area_flat_roof)

# calculate the peak capacity of this system in kWp
system_peak_capacity = panel_count * panel_peak_power / 1000

print(f"Based on the specified system characteristics, {panel_count} panels can be installed on a {roof_area} m² flat roof. \nThis corresponds to a total system capacity of {system_peak_capacity} kWp.")

Finally, we can estimate the total annual production of this 74.34 kWp system. For this estimate, we are simplifying by assuming that the total system production will scale linearly with the number of panels. In reality, system-level effects such as shading from nearby objects, soiling, and thermal effects due to panel arrangement can impact performance. However, for the sake of this tutorial, this simplification is acceptable. We will also divide the result by 1000 to convert the results from Wh to kWh.

# Calculate the monthly production of the entire PV installation 
system_production = panel_count * module_energy / 1000
monthly_production = system_production.resample('M').sum()

# Plot monthly production
fig = go.Figure()
fig.add_trace(go.Bar(x=monthly_production.index, y=monthly_production))
fig.update_layout(yaxis_title='Energy Produced (kWh)')
fig.show()

It’s quite interesting to see how, despite the fact that the single panel’s hourly peak production is only slightly different across the seasons, the total monthly production varies significantly (more than doubling from January to June). This implies that the increased system production during summer is not strictly due to higher production at peak times, but rather to having more hours of daylight each day, leading to longer periods of energy production. The effect of having more hours of daylight is more significant than differences in peak panel production between winter and summer.

Conclusion

In this tutorial, we demonstrated how to simulate the annual energy production of a photovoltaic system using just a few lines of Python code and a weather dataset. In the upcoming episodes of this series, we will explore how to align the estimated production with a building’s electricity consumption timeseries to determine the optimal system capacity. Additionally, we will compare the pvlib methodology with other methods used to simulate the operation of a photovoltaic system.