A Hybrid VMD–GWO–BiLSTM–Attention Framework for Accurate State-of-Health Prediction of Lithium-Ion Batteries Under Complex Degradation Dynamics

Abstract

Accurate State-of-Health (SOH) prediction for lithium-ion batteries is essential for the safe and stable operation of Battery Management Systems (BMSs). However, achieving high-precision estimation remains challenging due to the complex interplay of non-stationary degradation characteristics, observation noise, and bidirectional temporal dependencies. This study proposes a robust hybrid temporal prediction framework integrating Variational Mode Decomposition (VMD), Bidirectional Long Short-Term Memory (BiLSTM) networks, an Attention mechanism, and the Grey Wolf Optimizer (GWO). First, VMD is employed to decompose the raw SOH time-series signals, isolating observation noise and mitigating temporal non-stationarity. A BiLSTM network is then utilized to extract bidirectional temporal correlations throughout the aging process, while an attention mechanism adaptively allocates weights to focus on critical degradation features. Furthermore, GWO automatically searches for optimal network hyperparameters, eliminating the subjective bias of manual tuning and enhancing cross-condition generalization. The proposed framework is rigorously validated using full-lifecycle datasets of four lithium-ion batteries (CS2_35 to CS2_38) from the CALCE laboratory. Standard temporal networks (LSTM, GRU, and Transformer) serve as comparative baselines, alongside an intra-battery ablation study to quantify the performance gain of each module. Experimental results demonstrate that the proposed VMD–GWO–BiLSTM–Attention framework achieves enhanced prediction accuracy compared to standard baseline temporal models (LSTM, GRU, and Transformer) across all evaluated degradation profiles. For the CS2_35 battery with a smooth aging trend, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are minimized to 0.008 and 1.442%, respectively. Even under the severe capacity fluctuations of the CS2_38 battery, the model successfully controls the RMSE at 0.012 and the MAPE at 2.541%, exhibiting a significant error reduction compared to all baselines. These findings confirm that the synergistic integration of these four modules effectively addresses complex, nonlinear battery aging scenarios, offering a viable methodological baseline for deep-learning-based SOH prediction.