Forecasting-aided Robust Distribution State Estimation Based on Physics-aware Graphical Learning and Gaussian Process-aided Residual Modeling

A robust probabilistic distribution system state estimation method based on physics-aware graph network and residual learning method is proposed in this paper. Nodal load forecasting model is first developed to improve the observability of the system, followed by a graph attention network to capture structural correlation from the graph-structured data and identify the importance of neighboring nodes according to the forecasting errors of pseudo-measurement. After that, a residual learning module based on a Gaussian process with a composite kernel is employed to capture the behavior of the forecasting and estimation models. The embedding of uncertainties of pseudo-measurements and topology knowledge allows the proposed approach to achieve robustness to abnormal measurements. The residual modeling process further enhances the estimation performance of the proposed approach while providing uncertainty quantification of the state estimation results. Simulation results on standard IEEE test system illustrate the robustness of the proposed DSSE approach.