A novel multi-source spatial-temporal forecasting network for power prediction of electric vehicle charging stations

• MSSTON is proposed for EV charging load forecasting. • STN embeds CNN into Bi-LSTM to fuse spatial-temporal data. • MSSTON achieves a superior RMSE of 0.7705 on Boulder dataset. Accurate power demand prediction for Electric Vehicle (EV) charging stations is critical for smart grid stability. However, most existing methods focus predominantly on the temporal dependencies of single-station historical data, often neglecting the complex spatial correlations between different stations. To address this, this paper proposes a novel Multi-Source Spatial-Temporal Forecasting Network (MSSTON) for high-precision EV charging load forecasting. First, we design a Spatial-Temporal Network (STN) by embedding Convolutional Neural Network (CNN) blocks directly into Bi-directional Long Short-Term Memory (Bi-LSTM) units. This architecture facilitates the deep fusion of local spatial features and long-term temporal dependencies. Second, to fully leverage multi-source data, a Multi-Source Attention Mechanism (MSAM) is introduced. This mechanism dynamically weighs the importance of diverse data sources, effectively filtering noise and enhancing the extraction of high-correlation spatial features. Validated on the Boulder EVCS dataset, experimental results demonstrate that MSSTON achieves superior predictive performance with an root mean squared error (RMSE) of 0.7705 and an R-squared ( R 2 ) of 0.9868. The proposed method significantly outperforms traditional LSTM and hybrid CNN-BiLSTM baselines, exhibiting exceptional robustness and generalization ability across different geographical locations.