Bridging the Gap: A Real-World Analysis of AI Applications in Energy
In the first post of this series, we explored the potential impact of AI on the energy sector, with a particular focus on its application within the buildings industry. Following that, we went on to introduce the technical foundations of using machine learning models for energy consumption prediction.
In today’s issue, armed with a deeper understanding of how machine learning models are used to predict energy consumption, we will revisit the six use-cases introduced in the first post of the series. This should provide a clear overview of how energy prediction models are applied in each scenario. As a bonus, I’ll also mention the names of some of the leading companies working in each of these use-cases (no sponsorship involved).
Detecting Energy Waste and Consumption Anomalies: To effectively detect anomalies in energy consumption, the models used for this use-case must be able to capture complex patterns and unusual deviations. Artificial Neural Networks can be useful here, given their ability to model intricate, non-linear relationships. Additionally, tree-based models like Random Forests and Gradient Boosting Machines can also be effectively used for this purpose due to their capability to handle sensitivity to outliers. The Ento platform has AI-based digital advisors aimed at detecting unusually high consumption and reducing energy waste. Spacewell Energy, and Verdigris are other leading companies that harness advanced machine learning to swiftly detect and counteract energy consumption anomalies in buildings.
Evaluating Energy Efficiency Measures in Buildings through Advanced Measurement and Verification: Computational models used to verify energy savings usually operate on the level of individual buildings and work best when the data available has daily frequency or higher. Important independent variables commonly used in these models include weather factors, calendar features, and special events such as holidays or unusual occurrences like the COVID-19 pandemic.
While advanced Machine Learning algorithms can be used, many times there’s requirements to use clearly interpretable methodologies such as multivariate linear regression to estimate savings. This is because many times energy efficiency savings have to be agreed upon between different parties. The International Performance Measurement and Verification Protocol (IPMVP) provides clear guidelines regarding the use of statistical and machine learning models for the Measurement and Verification of energy efficiency projects. Ento provides an advanced M&V tool on its platform that follows the IPMVP guidelines. WattCarbon uses an open-source methodology to verify energy savings within the decarbonization portfolios offered on their marketplace. If you’d like to gain a deeper technical understanding of Measurement & Verification and how Machine Learning models are employed for this use-case, this review paper is a good place to start.
Optimizing Heating, Ventilation, and Air Conditioning (HVAC) Control using Model Predictive Controllers: Given the complexity of controlling HVAC systems based on multiple concurring factors such as indoor comfort levels, weather, and dynamic grid conditions, more complex models are usually preferred for this use-case. Model Predictive Control is an area where Neural Networks shine, but Gradient Boosting Machines also find some applications. These models are especially suited for grasping the intricate, almost real-time, connections between a range of independent variables and the energy demand of a building. The ideal frequency of the data here is often 5-minute or 15-minute intervals, allowing for a detailed picture of how the building's systems are performing.
The real challenge and opportunity lie in the great number of variables at play. These might range from indoor air quality, HVAC setpoint temperatures, and electric vehicle (EV) charging, to photovoltaic production, battery storage, and the conditions of the electrical grid. The goal is to harness this data and weave it into a cohesive model that doesn't just respond but anticipates and optimizes. Although these applications are at the cutting edge of research, there’s already many companies that are working on this kind of products. Examples are large established companies such as Siemens, but also new players in the field like Brainbox AI.
Estimating the Impact of Demand Side Flexibility Deployment: Demand Side Flexibility events are usually short-lived, demanding high-resolution data for accuracy. While hourly data might suffice for certain events, most require finer granularity, like every 15 or 5 minutes. These models incorporate variables similar to those in advanced Measurement & Verification, in addition they might also consider industrial production data, especially for facilities of large size. For this use-case, interpretability is key since compensations are usually paid based on the results of the model. For this reason, companies operating in this space often decide to work with open-source, interpretable, methodologies such as the ones implemented in OpenEEMeter. Companies like Recurve employ advanced M&V methods to quantify DSF deployment outcomes, ensuring precise rewards for their clients. Meanwhile, in Spain, Bamboo Energy provides and maintains a platform to support retailers and aggregators in the management of flexibility resources.
Supporting Energy Utilities in Day-Ahead Market Planning: In this use-case, an additional challenge is provided by the uncertainty around explanatory variables, like the upcoming weather forecast. Inputs frequently used in these models include weather conditions and calendar features, but to enhance precision, models can also directly incorporate historical energy demand patterns in the form of autoregressive terms. Given the critical nature of accurate forecasts, sophisticated models such as Gradient Boosting Machines (GBMs) or Artificial Neural Networks (ANNs) are often favored over more interpretable models. While specialized companies like Aleasoft, and Bidgely provide solutions in this realm, many large utility providers also craft their proprietary models for large-scale energy demand forecasting.
Characterizing Building Envelope Through Inference Models: The goal for this use-case is to extrapolate building characteristics, such as the heat loss coefficient of a building envelope, by analysing its energy consumption. Since the focus here is on inference, interpretable models like Linear Regression models or Generalized Additive Models are usually a suitable choice. The main explanatory features used are usually weather related. Indoor temperature readings or ventilation schedules when available can also be included in the models. The data frequency required is usually daily, although hourly data will allow for more accurate results. While building envelope analyses are of great importance when planning deep retrofits, these studies are more prevalent in academic research than mainstream commercial applications.
Conclusion
We started this series of posts with a question: how can Artificial Intelligence be used to reduce emissions in the buildings sector? To answer, we touched upon the fundamentals of employing machine learning models for energy prediction in buildings, discussing the core use-cases and analysing the mechanics of various machine learning algorithms. We then bridged these pieces of knowledge to uncover how specific algorithms are tailored for unique tasks in the field.
This concludes our first series of posts focusing on the role of Artificial Intelligence within the energy sector. I hope that the insights provided were clear and engaging, and that I managed to somehow spark your enthusiasm for this dynamic field.
For the upcoming issues, I have several ideas in store and would greatly appreciate your input on the types of content that capture your interest the most. Some ideas that I’ve been working on are:
High-level (non-technical) analyses detailing how AI is transforming the energy sector
Technical deep-dives, with code examples, showcasing AI applications on energy data
Regular updates on the latest developments in AI within the energy industry, with both a commercial and academic research perspective
P.S.
The list of organizations mentioned in this post is by no means exhaustive. For obvious reasons I could not enumerate all the companies working at the intersection of AI and energy, hence the focus was on those that I’m the most familiar with. If you work in a company in this sector that wasn’t mentioned, I would be happy to connect and discuss about your work. Feel free to reach out directly on LinkedIn.