A Short-Term Load Forecasting Method for Traction Substations Based on Physical Information Collaboration and Spatiotemporal Correlation

Accurate short-term traction load forecasting is crucial for optimizing railway operations. However, the strong fluctuations in high-speed railway loads and the general neglect of the physical relationships between adjacent substations in existing studies pose significant challenges to reliable short-term forecasting. To address these issues, this paper proposes a Lag-Adaptive Gradient Aware Network (LAGA-Net). Unlike isolated forecasting methods, LAGA-Net explicitly combines the physical information of train motion with deep learning methods to achieve collaborative load forecasting between adjacent traction substations (TSs). Specifically, it first calculates the cross-correlation coefficients of the load curves of adjacent TSs to quantify the train lag process and achieve load time-series alignment, effectively utilizing the historical load of upstream substations as prior information for load forecasting at this station. Based on this, a dual-stream gradient sensing encoder is proposed to capture the load amplitude and high-frequency pulses of the two TSs, improving the prediction accuracy of the model in highly volatile scenarios. Finally, an adaptive cross-attention mechanism based on Gaussian masks is designed to achieve spatiotemporal coupling and collaborative forecasting of the loads of two adjacent TSs using the aligned load representation information. Extensive experiments on real adjacent traction substation datasets demonstrate that LAGA-Net significantly outperforms existing state-of-the-art benchmark methods in terms of multi-step prediction and peak prediction accuracy, and exhibits strong robustness to operational uncertainties.