An Improved Transformer—Long and Short-Term Memory Network Modeling Strategy for Accurate State of Health Estimation of Lithium-ion Batteries in Power Stations

The growing concerns over environmental pollution and energy sustainability have accelerated the development of alternative energy technologies. Ensuring the reliable operation and safety of lithium-ion batteries requires a precise assessment of their health status. In this study, publicly accessible datasets from NASA and the University of Maryland serve as the foundation for analysis. Health-related features—such as peak points and their positions—are derived from the batteries’ capacity increment curves. Kalman filtering is utilized during data preprocessing to suppress measurement noise. To enhance state-of-health (SOH) prediction accuracy, a new hybrid model combining Transformer and LSTM architectures is introduced. Its performance is benchmarked against individual Transformer and LSTM models. The results reveal that the proposed approach offers superior predictive accuracy and robustness, achieving MAE, MSE, RMSE, and R2 values of 0.0013, 0.0002, 0.0021, and 0.9821, respectively, demonstrating a clear advantage over the standalone models.