Physics-informed Machine Learning for Power Systems

Contemporary advances in machine learning are increasingly contributing to the transition of power systems to a truly sustainable, resilient, and distributed infrastructure. Many challenges in power systems are governed by physics that are missing in general learning methods. For example, energy resource limits or security constraints cannot be addressed by methodological advances, while purely model-free approaches may require a huge number of data samples for defining all possible scenarios. As a result, a blind application of the machine learning toolset to solve stylized power-systems problems can inevitably suffer from a high sampling or computation complexity, and lack performance guarantees or interpretability. This panel puts forth some recent and exciting ideas on how to integrate power system domain expertise into the design, development, and improvement of learning-based computation solutions for data generation, modeling, inference, to decision making.