Machine Learning Applications In Power Systems

Machine learning (ML) applications have seen tremendous adoption in power system research and applications. For instance, supervised/unsupervised learning-based load forecasting and fault detection are classic ML topics that have been well studied. Recently, reinforcement learning-based voltage control, distribution analysis, etc., are also gaining popularity. Compared to conventional mathematical methods, ML methods have the following advantages: (i). better robustness against different system configurations due to its data-driven nature; (ii). better adaption to system uncertainties; (iii). less dependent on the modeling accuracy and validity of assumptions. However, due to the unique physics of the power grid, many problems cannot be directly solved using ML algorithms. Therefore, significant efforts are required to improve the adaptabilities of ML methods for power system applications. To this end, this thesis aims to fully investigate the effectiveness of ML in power systems and provide a different viewpoint towards ML. Novel ML methods are proposed in this thesis to solve a series of problems that are for the first time being solved using ML techniques. In this thesis, I extend the ML applications in power system by: (i). innovating the existing ML models to fit them into power system problems. (ii). reformulate the power system problems to make them solvable by ML techniques. Specifically, I integrate the human behavior analysis into Factorial Hidden Markov Model to conduct energy disaggregation for the individual household; combine the transfer learning, unsupervised learning, and supervised learning to form a framework for residential baseline load estimation; automate the transmission-level load modeling through reinforcement learning-based parameter estimation; apply reinforcement learning in solving large-scale Markov optimization for EV-based ancillary services; and combine the density-based spatial clustering of applications with noise (DBSCAN) with the conventional least square estimator (LSE), and enable the LSE to detect and recover the anomaly in the PMU measurements. The thesis demonstrates that by improving the existing ML models and reformulating power system problems we can not only expand the ML applications in power systems but also achieve better performance compared with the conventional methods