An adaptive spatiotemporal graph framework for multi-station electric vehicle charging availability forecasting

Accurate forecasting of available charging piles at public stations is vital for alleviating electric vehicle (EV) users’ range anxiety and improving the operational efficiency of urban charging infrastructure. However, complex spatiotemporal dynamics—such as time-varying spatial dependencies and heterogeneous station behaviors—pose significant challenges. To address this, a novel deep learning framework, MFA-STGCN (Multidimensional Feature-Enhanced and Adaptively Augmented Spatiotemporal Graph Convolutional Network), is proposed. It integrates: (1) a Data-Adaptive Graph Generation (DAGG) module to model evolving inter-station correlations; and (2) a Node-Adaptive Parameter Learning (NAPL) module to capture station-specific operational patterns. Experiments on a real-world dataset from Shenzhen, China, show that MFA-STGCN reduces MAE, RMSE, and MAPE by 31.6%, 24.7%, and 14.2% respectively, outperforming state-of-the-art baselines. Ablation studies confirm that DAGG and NAPL contribute complementary gains. This work provides data-driven support for real-time charging network management and offers methodological insights for modeling complex urban infrastructure.