A two-stage prediction method for power system cascading failures integrating static physical priors and autoregressive training

This study investigates three key challenges in forecasting cascading failures in power systems, namely effectively integrating multi-source physical priors with data-driven representations, inadequate discrimination among complex line categories, and the bias resulting from training-inference exposure. To tackle these obstacles, a two-stage prediction approach that combines static physical priors with autoregressive training is proposed. Initially, the method evaluates fault propagation risk. Subsequently, it predicts line failures and inter-event time intervals for confirmed cascading samples in a sequential fashion using autoregressive techniques. A three-layer hierarchical framework is devised for the proposed method, comprising input representation, training loss, and training strategy. Multi-source static physical prior encoding captures physical features in topology, line parameters, power flow, and load dimensions, integrated into a Transformer at the input layer. A confusion-aware hard-category reweighting mechanism adjusts class weights in the Focal loss based on misclassification rates calculated from a calibration set. Exponential scheduled sampling facilitates a seamless transition from reliance on ground-truth historical inputs to predicted inputs. The proposed method, evaluated on datasets from IEEE 118-bus and IEEE 300-bus systems, outperforms the Neural Hawkes Process (NHP) by 47.92 %age points on average and the Graph Convolutional Network (GCN) by 10.58 %age points in average sequence accuracy under three-step autoregressive prediction (AR-3 MSA). On the IEEE 118-bus and 300-bus test systems, the RMSE for time-interval predictions is significantly lower than that of NHP, with the E4 RMSE accounting for only 19.6%–38.7% and 10.4%–17.4% of the NHP RMSE, respectively. The three designs target heterogeneous fusion, hard category learning, and training-inference mismatch, leading to consistent and significant performance improvements across various system scales and sample conditions.