High-dimensional dynamic coupled graph spectral physics-informed neural network for power system prediction

Abstract

The reliable operation of power transmission and transformation systems faces unprecedented challenges due to the large-scale integration of renewable energy and increasingly complex power loads. While traditional data-driven models have shown potential, they often act as "black boxes" lacking physical interpretability. Physics-Informed Neural Networks (PINNs) offer a solution by embedding physical laws into the learning process, yet they struggle with the dynamic topologies and multi-physics coupling inherent in modern power grids. To address these challenges, we propose a novel framework termed as High-Dimensional Dynamic coupled Graph Spectral Physics-Informed Neural Network (HDGSPINN). The proposed method introduces a Spatiotemporal Graph Spectral Attention (STGSA) mechanism to capture complex spatiotemporal dependencies. Furthermore, a Hierarchical Multi-Physics Deep Coupling (HMPDC) module unifies electrical dynamics, 3D high-fidelity thermal fields, and physical aging models. To ensure training stability across diverse physical constraints, we design an advanced Gradient Dynamic Balance and Conflict Resolution (GDB-CR) loss strategy. Comprehensive experiments on the PowerGraph benchmark dataset demonstrate that HDGSPINN achieves state-of-the-art performance across both node-level and graph-level prediction tasks.