Multi-Scale Spatial-Temporal Graph Attention Network for Charging Station Load Prediction

The precision of electric vehicle (EV) charging station load forecasting is paramount for the efficient management of smart grids, fulfilling the charging demands of EV users and enhancing power system operations and planning. While existing forecasting methods have made strides, they often fall short in addressing the challenges of low spatial resolution and spatial heterogeneity. Low spatial resolution, characterized by sparse data distribution and a dearth of information between geographically diverse charging stations, coupled with spatial heterogeneity, which is evident in the fluctuating energy demand across different regions, requires a sophisticated approach. In this study, we introduce the Multi-Scale Spatial-Temporal Graph Attention Network (MSSTGAN), a pioneering framework that adeptly navigates these complexities. MSSTGAN employs an encoder-decoder architecture underpinned by split-pyramid attention to resolve issues of low spatial resolution, enabling the model to process multi-scale data and capture temporal dynamics effectively. Moreover, it incorporates graph attention mechanisms to manage spatial heterogeneity in energy demand, ensuring a nuanced understanding of regional variations. Our framework’s efficacy is underscored by its superior performance over existing baselines on four real-world EV charging station datasets, showcasing its prowess in tackling the intricacies of low spatial resolution and spatial heterogeneity.