Gravitational Search Algorithm Based LSTM Deep Neural Network for Battery Capacity and Remaining Useful Life Prediction With Uncertainty

An accurate estimation of the remaining useful life (RUL) and capacity of lithium-ion batteries (LIBs) can guarantee safe and reliable operation and help to make wise replacement decisions. This paper presents an improved approach for predicting the RUL and capacity of LIB using a long short-term memory (LSTM) deep neural network-integrated with a gravitational search algorithm (GSA) to address the challenges associated with predicting battery life. Initially, data cleaning is carried out to minimize any negative impacts that can reduce the convergence rate. Abnormal data are replaced with highly correlated data, and the data is standardized. Moreover, the LSTM model hyperparameters are optimized using the GSA optimization technique. To evaluate the robustness of the proposed method, 15 prediction samples are generated to calculate the uncertainty levels (95% CI) of the predicted RUL. The proposed method is assessed using aging data from the NASA battery dataset. Its performance is compared with baseline LSTM, baseline GRU, BiLSTM, and LSTM-based particle swarm optimization (PSO) models across various error metrics. The robustness of the proposed method is verified by benchmarking it against other existing approaches for predicting RUL and capacity. The results indicate that the LSTM-GSA model outperforms in prediction accuracy, achieving RMSE values of 1.04%, 1.15%, 1.26%, and 0.92% across different battery cases at both early and later cycle stages. Overall, this research provides a promising solution for predicting RUL and the future capacity of LIBs with uncertainty, which is essential for ensuring the safe and efficient operation of energy storage systems.