Learning the Unobservable: High-Resolution State Estimation via Deep Learning

Abstract

The problem of achieving the fast-timescale state estimation for a power system with limited deployment of phasor measurement units is considered. A deep neural-network architecture that integrates bad-data detection, data cleansing, and the minimum mean squared error state estimation is developed. It includes a universal bad-data detection and a Bayesian state estimation subnetworks. A novel universal bad-data detection technique is proposed that requires no knowledge about data distributions under regular and irregular operating conditions. The subnetwork for universal bad-data detection consists of an inverse generative model and a coincidence test. It is implemented through the training of a generative adversary network and an auto-encoder using slow-timescale historical data. The Bayesian state estimation subnetwork is trained through a generative adversary network with embedded physical models of the power system. Comparing with the conventional weighted least squares approach to state estimation, the proposed minimum mean-squared error state estimator does not require observability. Simulations demonstrate orders of magnitude improvement in estimation accuracy and online computation costs of/ver the state-of-the-art solutions.