An electric vehicle charging demand prediction approach based on a Graph-based Spatio-temporal Attention Network

The accelerated adoption of electric vehicles (EVs) is fundamentally reshaping urban transportation and energy systems. However, the growing charging demand creates significant stress on urban power grids, particularly during peak hours in megacities like Shenzhen. Accurate short-term prediction of charging demand is crucial for optimizing infrastructure layout, maintaining grid stability, and supporting data-driven energy policies for sustainable urban development. This study proposes a novel Graph-based Spatio-Temporal Attention Network (G-STAN) that integrates graph convolutional networks and attention mechanisms to address the dynamic spatio-temporal characteristics of EV charging demand. The model employs a Residual Temporal Convolution Network (Res-TCN) to capture short-term load fluctuations, a Simple Graph Convolution Attention (Sim-GCA) module to model spatial interactions across 247 traffic zones, and a Temporal Pattern Attention (TPA) module to focus on peak hours and key functional areas. Evaluated on a real-world citywide charging dataset, G-STAN outperforms existing models by improving RMSE, MAE, and MAPE by 21.92 %, 36.96 %, and 16.92 %, respectively. With a lightweight design and multimodal input integration, the proposed framework enables efficient, scalable, and policy-responsive forecasting. This study proposes a novel prediction paradigm that not only significantly improves the prediction accuracy compared to state-of-the-art models, but also enhances model interpretability, scalability, and responsiveness to policy signals. It provides an actionable framework for real-time intelligent regulation of EV charging behaviors, supporting the synergy between urban energy management and sustainability goals in smart city ecosystems.