A Tensor Decomposition Personalized Federated Learning Method for Distribution System State Estimation Considering Dynamic Impacts of Transmission Network

Traditional methods for distribution system state estimation are inadequate in capturing and coordinating the dynamic effects imposed by the transmission system. In addition, data privacy breach risks in power systems have become increasingly pronounced. To address these challenges, this paper proposes a Tensor Decomposition Personalized Federated Learning (TDPFed)-based Dynamic-Static Integrated State Estimation (DSISE) method for the coordinated transmission-distribution (T&D) networks. A DSISE model is constructed where the Extended Kalman Filter (EKF)-based dynamic state estimation is for the transmission network, and the Generalized Maximum Likelihood approach based on Projection Statistics (GM-PS) for static state estimation is for the multi-area three-phase unbalanced distribution network. To reduce model communication costs and improve convergence speed, TDPFed designed a two-layer loss function to realize the decoupling between the personalized model and the tensor local model by controlling the distance. In addition, this paper also designs an efficient distributed learning strategy and model aggregation methods for TDPFed. Through theoretical analysis and experimental validation, it has been proven that the TDPFed-DSISE framework not only reduces communication costs but also significantly improves estimation accuracy compared to traditional federated learning algorithms.