State-of-health prediction of lithium-ion batteries using feature fusion and a hybrid neural network model

With their high energy density and long cycle life, lithium-ion batteries are vital components of energy storage systems. However, accurate State of Health (SOH) prediction remains challenging due to nonlinear degradation over prolonged cycling. To address this issue, a hybrid neural network model based on feature fusion is proposed. Capacity-Voltage (Q_V), Time-Voltage (T_V), and incremental capacity (dQ/dV_V) features, along with a fused three-dimensional composite feature representation, are utilized to comprehensively characterize battery aging dynamics. Feature extraction is performed using four independent Temporal Convolutional Networks (TCNs), followed by an attention mechanism to adaptively weigh feature importance. To capture temporal dependencies, a Bidirectional Gated Recurrent Unit (BIGRU) is employed, significantly improving prediction accuracy. Furthermore, the Beluga Whale Optimization (BWO) algorithm is applied to optimize model parameters, ensuring enhanced predictive performance. The proposed approach achieves a MAPE of 0.2573 % and a RMSE of 0.3173 % on the MIT dataset. To evaluate its cross-material generalizability, a transfer learning strategy with fine-tuning is introduced. Evaluation on the Oxford battery aging dataset resulted in a MAPE of 0.1954 % and an RMSE of 0.1979 %, demonstrating the model's robustness across different datasets. The proposed approach outperforms conventional methods in both accuracy and adaptability, offering a robust and scalable solution for SOH estimation across diverse battery datasets.