Explainable MFDWA-GRU Model for Battery State of Health Estimation

The operational safety of lithium-ion batteries is heavily dependent on accurate battery state of health (SOH) assessments. This paper introduces a novel hybrid algorithm: a multi-faceted temporal convolutional network with dynamic weight adaptation (MFDWA) combined with a gated recurrent unit (GRU) to estimate the SOH of lithium-ion batteries across various experimental setups. The MFDWA architecture includes multiple temporal convolutional network (TCN) branches, each containing a dilated TCN block, attention blocks, and dynamic weight adaptation components. This multi-branch structure captures multiscale patterns through different dilation rates, while attention blocks select essential input features. The dynamic weight adaptation mechanism further enhances performance by addressing evolving battery dynamics. The GRU component tracks temporal information sequences by retaining relevant past states. The proposed hybrid algorithm performs better than the existing state-of-the-art methods, achieving lower mean absolute error (MAE) and root mean square error (RMSE) values across multiple datasets while maintaining parameter compatibility. These results indicate the model's robustness across different initial conditions, battery combinations, and varying prediction horizons. Additionally, the integration of explainable artificial intelligence (XAI) through the SHapley Additive exPlanations (SHAP) technique in the proposed hybrid model prioritizes input features, improves model performance, and enhances interpretability, thus fostering user trust.