Interpretable State Estimation in Power Systems Based on the Kolmogorov–Arnold Networks

Abstract

Listen

Translate article

Power system state estimation is a critical task for ensuring stable grid operation and serves as the foundation of grid control and analysis. Conventional approaches largely involve field measurements, network topology, and manual anomaly detection, which present significant limitations, particularly while dealing with dynamic and complex power systems. In recent years, deep learning techniques have been progressively applied in this field to overcome the shortcomings of conventional approaches, which are based on mathematical models and static analysis. However, existing deep learning techniques primarily focus on power system security analysis and computational resource management. In spite of the powerful capabilities in supervised learning tasks, the lack of interpretability still makes deep learning models less convincing, and thus might hinder their practical applications. In response to this issue, we apply a computational model for power system state estimation based on the Kolmogorov–Arnold network (KAN) model with learnable activation functions, visualization capabilities, and pruning features. From the perspective of feature interpretability, we find the influence of bus features on the model output, such as bus voltage magnitude. Moreover, through analysis of the internal structure of the model, we uncover a possibility of potential mechanisms of power system state estimation. Experimental results show that our study not only enhances the interpretability of power system state estimation but also effectively ensures grid security and stability through state estimation.