State of Health Prediction of Pure Electric Vehicle Batteries Based on Patch Wavelet Transformer

The accuracy of onboard power battery capacity prediction is often limited due to excessive reliance on laboratory data and the neglect of complex usage environments. To address this issue, a Transformer-based model, named PWT, which integrates a patching strategy with wavelet decomposition, is proposed. By utilizing multi-scale temporal feature extraction and attention mechanisms, the model effectively enhances the capability of battery degradation modeling. To tackle the challenge of limited capacity label availability, a fuzzy Kalman filtering model based on ampere-hour integration is designed, reducing the relative error in SOH estimation to 0.906% and significantly improving label accuracy. Furthermore, a charging behavior scoring mechanism based on fuzzy membership functions and a current–temperature interaction feature matrix is constructed to enhance the model’s sensitivity to degradation factors. Experimental results show that the proposed method outperforms LSTM, Transformer, and PatchTST under various real-world operating conditions, achieving a worst-case RMSE of 0.0226, MAPE of 0.0725, and R 2 of 0.903, demonstrating higher accuracy, robustness, and computational efficiency. In conclusion, the proposed method exhibits promising prospects in both theoretical research and engineering applications, providing a novel and effective approach to power battery health management.