Machine learning-driven time series analysis for SOH prediction of lithium-ion batteries

Energy storage batteries are essential for stabilizing renewable energy systems and improving power grid efficiency. However, challenges such as capacity degradation, limited data quality, and the need for real-time evaluation highlight the importance of accurate State of Health (SOH) prediction. This study evaluates the effectiveness of Random Forest, Long Short-Term Memory (LSTM), and Bidirectional Long Short-Term Memory (Bi-LSTM) models in predicting SOH across single lithium-ion cells, cells under environmental influence, and battery modules. The Bi-LSTM model achieved a Mean Absolute Error of 0.00668, a Mean Squared Error of 0.0042, and an R² value of 0.9253 in single-cell prediction. In comparison, the Random Forest model recorded a Mean Absolute Error of 0.0523, a Mean Squared Error of 0.0159 and an R² of 0.8960, indicating a reduction in error of over 69% and a significant improvement in predictive accuracy. Incorporating physically meaningful features such as discharge time and plateau voltage further enhanced model performance. These results demonstrate the Bi-LSTM model's strong ability to capture long-term temporal dependencies and its potential for improving intelligent battery health monitoring in real-world energy storage systems.