Sparrow Search Algorithm-optimized CNN–BiLSTM for lithium-ion battery capacity degradation prediction with SHAP-based interpretability

Accurate prediction of lithium-ion battery capacity degradation is essential for battery health assessment and maintenance planning. This study presents an interpretable capacity prediction framework integrating a convolutional neural network (CNN), bidirectional long short-term memory (BiLSTM), Sparrow Search Algorithm (SSA)-based hyperparameter optimization, and SHapley Additive exPlanations (SHAP). Six health indicators derived from charging, discharging, and incremental-capacity profiles are used as model inputs. The CNN extracts local feature patterns, while the BiLSTM–LSTM–BiLSTM structure performs nonlinear feature processing. SSA optimizes the learning rate, regularization coefficient, and recurrent-layer hidden units, whereas SHAP quantifies the contributions of individual health indicators to the predicted capacity. The framework is evaluated on three experimentally aged 1.2 Ah lithium-ion batteries and the NASA B0005 dataset using chronological training, validation, and testing subsets. For the three experimental batteries, the reported RMSE values range from 0.003196 to 0.005025 Ah, with R² values between 0.922 and 0.989. On NASA B0005, the model achieves an RMSE of 0.0373 Ah and an R² of 0.8355. Under the adopted evaluation conditions, the proposed framework achieves lower prediction errors than four internal baseline models. SHAP analysis identifies the charging-current integral, time to minimum discharge voltage, and incremental-capacity peak as important contributors to the predicted capacity. These findings demonstrate the potential of combining capacity degradation prediction with model-specific feature interpretation to support battery health monitoring and condition-based maintenance decisions, while further cross-battery validation is required before practical deployment.