Load probabilistic forecasting for electric vehicle charging station cluster: Adaptive multivariate spatiotemporal graph fusion technology

The complexity, randomness, and spatial heterogeneity of electric vehicle user charging behaviors pose significant challenges for load forecasting. Existing literature primarily focuses on load prediction for individual charging stations, with relatively limited research at the charging station cluster level. Moreover, current methods fail to directly capture the power fluctuations caused by the dynamic changes in different load types and the number of connections. To address this, this paper proposes an Adaptive Multivariate Spatiotemporal Graph (AMSTG) model framework for probabilistic EV charging load forecasting. The model integrates a time Transformer with a multi-layer graph fusion module to capture the complex spatiotemporal dynamic coupling relationship between charging power and station occupancy, effectively adapting to the dynamic evolution driven by the composition and number of various load types. Furthermore, a Copula-based probabilistic decoder is employed to capture the joint distribution among multiple variables, enabling reliable uncertainty quantification and accurate imputation of missing values. The effectiveness of model is further demonstrated through experiments on real-world data collected from 247 public EV charging stations in Shenzhen, China. Compared with various baseline approaches, the proposed method achieves superior accuracy in forecasting loads across multiple stations.