Interpretable Spatio-Temporal Attention for Shared Charging Load Prediction

Abstract

Urban bus electrification and the proliferation of shared charging infrastructure have intensified the demand for accurate, interpretable load forecasting to support reliable grid dispatch and energy management. Existing spatio-temporal graph neural networks, however, rely on static geographic-distance adjacency matrices and fixed spatial-temporal fusion weighting, limiting their ability to capture the dynamic, dual-entity nature of scheduled bus fleets and private electric vehicle charging demand—while offering little operational transparency. To address these limitations, we propose the Spatio-Temporal Attention Network (ST-Attention), an interpretable forecasting framework built on three core components: a behavior-driven multi-layer adjacency matrix that encodes bus origin-destination flows, gravity-model private vehicle flows, and spatial proximity to reflect true traffic-induced correlations; a parallel spatial-temporal encoding framework combining graph attention and Transformer modules; and a gated fusion mechanism that dynamically weights spatial and temporal features at each node and time step. This design enables the model to adapt to shifting dominant factors throughout the day. Evaluated on a physics-grounded simulation dataset of shared charging stations, ST-Attention achieves a Mean Absolute Percentage Error of 53.45% and a Mean Absolute Error of 85.07 kW, with only a marginal accuracy gap relative to black-box baselines, while delivering full interpretability. The learned gating weights further reveal actionable operational patterns—explicitly indicating when spatial spillover versus local historical context drives each prediction—providing grid operators with transparent, reliable insights for infrastructure planning and safety-critical dispatch.