Physics-informed Machine Learning for Power Systems

The increasing deployment of sensors in power transmission and distribution systems around the world has generated an unprecedented amount of data. The traditional software tools and computing platforms used by electric utilities are not capable of effectively utilizing the big and heterogeneous data sets. Off-the-shelf machine learning algorithms have been widely adopted by researchers and developers to solve a myriad of problems in power systems. These initial efforts have achieved some great results in areas such as load and renewable energy forecasting and equipment predictive maintenance. However, when it comes to power system monitoring, sequential decision-making, optimization and controls problems, pure data-driven algorithms often do not yield satisfactory results. To fill the knowledge gap, this seminar covers the development of physics-informed machine learning algorithms by synergistically combining power system models and advanced machine learning techniques. The unique power system domain knowledge, information and models that have been integrated into machine learning algorithms include high/low entropy of certain power system sensor data, low-rank property of streaming data matrix, physical model for generation resources, power flow models, and power system dynamic and control models.