EVMamba: An Efficient Spatio-Temporal-Frequency State-Space Model for Electric Vehicle Charging Load Forecasting

As electric vehicles (EVs) become ubiquitous components of cyber-physical energy systems, their aggregated charging demand introduces unprecedented uncertainty and ramping events to power grids. Accurate, long-horizon load forecasting is, thus, a prerequisite for reliable grid dispatch and infrastructure planning. However, traditional data-driven models often falter when handling the complex nonlinearity and long-range dependencies inherent in large-scale EV charging data. To bridge this gap, we propose EVMamba, a novel forecasting framework that leverages a dual-stream encoder–decoder architecture to systematically capture these cross-domain dependencies. We design a unique wavelet-spatio-temporal embedding strategy to explicitly model the multifrequency characteristics of load profiles, effectively filtering noise while retaining essential trend and seasonal information. Crucially, we replace the computationally expensive self-attention mechanism with a dual-stream Mamba backbone. This architecture enables the model to capture bidirectional temporal contexts over extended horizons with linear complexity, making it highly suitable for resource-constrained industrial informatics scenarios. Comprehensive experiments on two datasets confirm that EVMamba achieves state-of-the-art performance, delivering significant improvement in forecasting precision. The proposed framework offers a viable pathway to improve the observability and resilience of next-generation smart grids under high EV penetration.