Risk assessment of insulator contamination in thermal power plants by integrating dynamic physical priors and spatiotemporal heterogeneous graphs

To address the problem of predicting insulator fouling risks at thermal power plants resulting from the combined effects of various source-side facilities and meteorological conditions, we propose the Causal-STHGN model. This spatiotemporal heterogeneous graph model integrates dynamic physical priors. This paper first constructs a semantic graph of power plant entities. Source-side facility nodes are mapped to insulator monitoring points as heterogeneous nodes. Then, by integrating spatial geometric relationships, wind field directionality, source-side state characteristics, and moisture condition modulation, the model calculates dynamic physical causal priors and edge-weight priors. Furthermore, it constrains spatial dependency learning through a prior-bias graph attention mechanism. A parallel temporal structure is employed to jointly model short-term disturbances and long-term evolutionary trends. Finally, leakage current is used as a supervised proxy to calibrate the continuous risk representation, thereby evaluating the model’s ability to characterize key processes such as moisture-induced excitation and rain scouring. Simulation results indicate that the proposed method demonstrates strong overall performance across metrics including overall prediction accuracy, risk alerts, and physical consistency, and exhibits risk responses that are more in line with reality within typical event windows.