A VMD-BiGRU-Attention-Based Method for Predicting the State of Charge of Rail Transit Lithium Batteries

With the increasing reliance of rail transit systems on new energy power supply systems, the accurate and reliable estimation of the state of charge (SOC) of lithium-ion batteries has emerged as a central concern in battery management system design. In this work, a hybrid SOC prediction model is developed by integrating Variational Mode Decomposition (VMD), Bidirectional Gated Recurrent Unit (BiGRU), and the Attention mechanism, aimed at SOC prediction for rail transit lithium batteries. First, VMD is employed to decompose the original voltage and current signals into multiple Intrinsic Mode Functions (IMFs), enabling multi-scale feature extraction, effective noise suppression, and enhanced signal predictability. Next, all IMFs are concatenated and reconstructed into a multi-modal input sequence, which is then fed into the BiGRU network to better capture both forward and backward temporal dependencies. Meanwhile, the attention mechanism is introduced to dynamically assign weights to different time steps and IMF components, thereby enhancing the model’s focus on critical features. Finally, the outputs of sub-networks are fused through weighted aggregation to achieve accurate SOC prediction. Experiments conducted on a publicly available lithium battery dataset demonstrate the effectiveness of the proposed VMD-BiGRU-Attention model. Compared with GRU, BiGRU, and VMD-GRU models, the proposed approach achieves superior prediction accuracy.