Accurately forecasting the load of electric vehicle charging stations (EVCSs) is crucial for optimizing grid operations and facilitating EV integration, yet existing methods struggle to capture the intricate spatio-temporal dependencies and the impact of influential EVCSs within charging networks. To address this, we propose a novel framework, Causality-Aware Dynamic Multi-Graph Convolutional Network (CADGN), a multi-graph convolutional network that integrates causal inference and critical node modeling. It consists of two core modules: the Causality-Aware Graph Learning Module (CAGLM) uncovers and represents causal relationships between EVCSs, while the Critical Relationship Graph Learning Module (CRGLM) dynamically models the evolving connections among critical EVCS nodes. Temporal patterns extracted from these modules are then fused to generate accurate load predictions. Extensive experiments using real-world datasets of hourly charging data from multiple cities demonstrate CADGN’s superiority over state-of-the-art EVCS load forecasting models, particularly for short-term and mid-term horizons. Notably, our model achieves an average 4.7% reduction in Mean Absolute Error (MAE) compared to Graph WaveNet across all datasets and prediction horizons, highlighting the practical benefits of considering both causal and critical relationships for enhanced grid operations and EV integration. These results emphasize the importance of incorporating causality and the identification of critical relationships in the EVCS load forecast to achieve higher accuracy.