Two-layer decomposition-fused hybrid deep learning enables data-driven electricity demand forecasting for battery swapping station

Accurate electricity demand forecasting for battery swapping stations (BSS) is essential for optimizing grid stability and operational efficiency. This study presents a dual-stage decomposition framework combining complementary ensemble empirical mode decomposition with adaptive noise and feature mode decomposition to address nonlinear and volatile demand patterns. A hybrid deep learning architecture integrating bidirectional temporal convolutional networks with gated recurrent units and attention mechanisms is developed, enhanced through automated hyperparameter optimization via a hippopotamus-inspired metaheuristic algorithm. Validated using real-world operational data from a BSS in February in China, the proposed model achieves maximum reductions in mean absolute error (MAE) of 22.87 % and 44.39 % for charging and swapping demand predictions compared to existing benchmarks. The results demonstrate that the integration of decomposition techniques, metaheuristic optimization, and bidirectional deep learning significantly improves prediction accuracy across 28-day, weekly, and daily horizons. This approach provides a robust foundation for demand response strategies, grid interaction planning, and sustainable BSS expansion.