With the increasing penetration of electric vehicles, charging facilities and load will become key growth areas for urban power grid development. Accurately forecasting the spatial distribution and temporal variation of electric vehicle charging load is crucial for ensuring the safe and stable operation of urban power grids. To address this, a forecasting method based on the stacked heterogeneous spatio-temporal graph network (SHSGN) model is proposed. This method improves the accuracy of electric vehicle load predictions by modeling the spatial distribution patterns of urban charging loads and the diverse charging behaviors of users. A two-layer heterogeneous graph neural network model is established to represent the charging station-level and regional-level load nodes. By enhancing the cross-regional load transfer patterns, this model improves the ability to capture spatial load characteristics. A multi-GRU parallel decoding neural network model is developed to accurately express the differentiated charging patterns of each charging station node through independent decoding tasks. Validation using real charging load data from a Chinese city demonstrates that the proposed method reduces the charging load prediction error to 6.77%, achieving a 2.12% improvement over traditional methods.