Automated and Robust System Discovery In Power Systems Assets Using Physics-Informed Machine Learning

System discovery, such as parameter estimation or model reconstruction, in complex power system assets can play crucial roles in condition monitoring, troubleshooting and root cause analysis. However, robust system discovery in presence of noisy or sparse data is a challenging problem that requires solving an inverse problem and quantifying uncertainties. In this paper we address this problem and propose an approach to find the solution by using Physics-Informed Machine Learning approach for modeling cause and effect in power systems assets with polynomial dynamics. In a probabilistic setting we establish the merits of the proposed approach using empirical results. In particular, we benchmark our proposed end-to-end method with a baseline that uses Bayesian Neural Network with dropout and show that our approach outperforms baseline by 79% of reduction of validation error, 34% reduction of training time and 97.5% reduction of inference time.