High-Resolution Real-Time Power Systems State Estimation: A Combined Physics-Embedded and Data-Driven Perspective

Abstract

Real-time perception of the power system operating state with high resolution is essential for enabling online dynamic security assessment. However, challenges associated with limited redundant measurements, dynamic model complexity and balancing state and non-state variables' accuracy hinder conventional model-driven and data-driven state estimation (SE) methods from delivering real-time states with high temporal-spatial precision. This paper proposes a novel physics-embedded and data-driven SE framework. By incorporating physics knowledge into both SE model development and training, this framework systematically bolsters previous high-resolution data-driven SE framework. By utilizing the physics model to translate hybrid measurements into node features and provide recent system state, the multi-head graph attention network is employed to extract spatial features, correcting discrepancies between the current and recent states through a Residual Network. To enhance accuracy of both state and non-state variables, the SE model undergoes training by a novel physics-embedded training method. This approach adaptively adjusts the weighting of state and non-state variables in the loss function, ultimately enhancing their estimation accuracy. Case studies verify its superior performance in terms of accuracy, efficiency, scalability and robustness on the IEEE 39-bus and 118-bus test systems. Furthermore, its advantages compared to traditional data-driven methods are proved theoretically in this paper.