A regional short-term load forecasting model based on a cross-attention causal spatio-temporal convolutional network

• Proposed a novel framework combining causal inference and graph convolution. • Introduced seasonal trend decomposition to remove spurious correlations. • Employed cross attention to integrate temporal and spatial features. • Validated on real multi-region load data with superior accuracy and efficiency. Accurate prediction of regional short-term load is crucial for optimizing power dispatch and renewable energy integration. Existing methods overlook the causal characteristics of load flow between different regions. This paper proposes a regional short-term load prediction model that combines Graph Convolutional Network (GCN) and Informer for spatio-temporal feature fusion. In terms of spatial feature extraction, the model first performs Seasonal-Trend Decomposition using Loess (STL) on regional load data to obtain high-frequency components, then conducts Fast Causal Inference (FCI) on these high-frequency components to derive a dynamic adjacency matrix that captures spatial causal features, and finally employs a two-layer Graph Convolutional Network (GCN) to extract spatial causal features. For temporal feature extraction, Informer and one-dimensional convolutional layers are used to capture load temporal features across different time scales. Finally, a cross-attention mechanism is used to dynamically integrate spatial and temporal features. Experiments were conducted using real load data from the New England region of the United States, and comparisons with other state-of-the-art models showed reductions of at least 5.34 MW in RMSE, 4.84 MW in MAE, and 0.3 % in MAPE. The experimental results validate the effectiveness of the proposed forecasting model.