Improving Lithium-Ion Battery Life Prediction using CEEMDAN Decomposition and Machine Learning

The remaining useful life (RUL) of lithium-ion batteries is a key indicator for the optimal design of battery management systems and ensuring equipment safety. Most existing RUL prediction methods retain only the trend term of the capacity curve, neglecting the potential information in high-frequency fluctuations. Therefore, this paper presents a combined model based on adaptive white noise complete integrated empirical mode decomposition (CEEMDAN), gradient boosted decision tree (GBDT), and long-short-term memory neural network (LSTM). CEEMDAN is employed to reduce noise and extract fluctuation components at different frequencies. GBDT assigns weights to each component to assess its contribution to the interpretation of the original data. The final prediction is obtained by weighted reconstruction of the LSTM prediction results for each component. The experimental results show that using the NASA lithium battery dataset, the CEEMDAN-GBDT-LSTM model significantly improves RUL prediction accuracy and outperforms single models in predictive performance.