Short-term regional EV charging load forecasting based on GAT and GRU with trip distribution estimation

Accurate forecasting of electric vehicle (EV) charging demand is crucial for developing coordinated charging strategies and reducing the negative impacts of large-scale, uncoordinated EV integration on power grid operations. Although previous studies have proposed predictive models using traffic simulations and machine learning with both dynamic and static features, most existing methods still fail to capture inter-nodal demand correlations, especially those involving long-distance nodes and heterogeneous data sources. To overcome these limitations, this study introduces STGR-Net, a novel short-term regional forecasting framework that integrates Graph Attention Networks (GAT), Gated Recurrent Units (GRU), and traffic distribution principles. The framework combines weekly aggregated charging data with fine-grained time series to extract spatio-temporal correlations, with particular focus on interactions among distant nodes. Based on these correlations, three matrices are constructed: (1) a dynamic gravity matrix from charging volume patterns, (2) a dynamic correlation matrix using inter-nodal Pearson coefficients, and (3) a static physical adjacency matrix. These components collectively form a novel triple-stream graph attention architecture. Temporal features are captured by parallel GRU encoders operating on three temporal sequences: recent, daily-periodic, and weekly-periodic. Each encoder is augmented with Reversible Instance Normalization (RevIN) to mitigate distributional shifts across different time periods. An adaptive gating mechanism further fuses temporal features with multi-dimensional spatial representations before prediction. Experiments on large-scale datasets from Beijing and Shenzhen show that STGR-Net achieves significant improvements in prediction accuracy over benchmark models. Ablation studies further confirm the contribution of the triple-stream graph architecture. The framework also shows strong practical utility for grid load management, charging service optimization, and infrastructure planning, supported by its efficient and accurate regional forecasting capability.