AI for Energy and Energy for AI

Our goal is to ensure that the energy sector can not only accommodate the rapid growth in electricity demand driven by AI technologies, but also harness AI itself to improve efficiency and reliability across energy systems and applications. We use learning-to-optimize methods to enhance the fidelity of energy management tools for power system operations by addressing challenging nonconvex problems (e.g., uncertainty in unit commitment, nonlinear power flows, pump/turbine scheduling). We also employ machine learning to improve control and decision-making in built environments—for example, how to optimize thermal management in residential buildings or large-scale data centers to achieve high performance with minimal energy losses. Our work in AI for energy spans foundational modeling and theory through to deployable tools, including contributions to NeuroMANCER, an open-source differentiable programming library. All these AI innovations, in turn, seek to improve power grid’s ability to serve the surging electric demand from AI more affordably. For example, the above mentioned ICARUS project leverages the learning-to-optimize framework for proxy rules to optimize data center placement.