The RUL Prediction of Li-Ion Batteries Based on Adaptive LSTM

Abstract

With the widespread adoption of electric vehicles and energy storage systems, predicting the remaining useful life (RUL) of lithium-ion batteries (LIBs) is critical for enhancing system reliability and enabling predictive maintenance. Traditional RUL prediction methods often exhibit reduced accuracy during the nonlinear aging stages of batteries and struggle to accommodate complex degradation processes. This paper introduces a novel adaptive long short-term memory (LSTM) approach that dynamically adjusts observation and prediction horizons to optimize predictive performance across various aging stages. The proposed method employs principal component analysis (PCA) for dimensionality reduction on publicly available NASA and Mendeley battery datasets to extract health indicators (HIs) and applies K-means clustering to segment the battery lifecycle into three aging stages (run-in, linear aging, and nonlinear aging), providing aging-stage-based input features for the model. Experimental results show that, in the NASA dataset, the adaptive LSTM reduces the MAE and RMSE by 0.042 and 0.043, respectively, compared to the CNN, demonstrating its effectiveness in mitigating error accumulation during the nonlinear aging stage. However, in the Mendeley dataset, the average prediction accuracy of the adaptive LSTM is slightly lower than that of the CNN and Transformer. These findings indicate that defining aging-stage-based adaptive observation and prediction horizons for LSTM can effectively enhance its performance in predicting battery RUL across the entire lifecycle.

Conflict of Interest Statement

The authors declare no conflicts of interest.