Towards enhanced remaining useful life prediction of lithium-ion batteries with uncertainty using optimized deep learning algorithm

An essential component of assessing lithium-ion battery (LIB) performance, reliability, and administration in the application of battery health monitoring and management is determining the battery's Remaining Useful Life (RUL). However, existing RUL prediction approaches have difficulties with variability and nonlinearity that occur during battery degradation, data extraction, feature extraction, hyperparameters optimization, and prediction model uncertainty. To address these problems, this paper introduces a novel hybrid approach for RUL prediction, combining a Lightning Search Algorithm (LSA) with a Long-Short Term Memory (LSTM) deep learning model. At first, the hybrid LSA + LSTM model is trained using a comprehensive framework comprising 31 data features, utilizing a mathematical systematic sampling (SS) approach. This sampling technique enables the identification of 10 related data features including temperature, voltage, and current, recorded during each charging cycle from the LIB parameters. Moreover, the LSA optimization technique is introduced to optimally determine the LSTM deep neural model hyperparameters including the number of hidden neurons, learn rate, epoch, learn rate drop factor, learn rate drop period, and gradient decay factor. The effectiveness of the proposed LSA + LSTM model is assessed using battery aging data from the NASA dataset. In addition, to validate the prediction performance of the proposed LSA + LSTM model, extensive comparisons are performed with other popular optimization-based deep learning methods including artificial bee colony (ABC) based LSTM (ABC + LSTM), gravitational search algorithm (GSA) based LSTM (GSA + LSTM), and particle swarm optimization (PSO) based LSTM (PSO + LSTM) model using different error matrices. The robustness of the proposed method is further validated with existing literature as well as with another battery dataset obtained from the MIT Stanford dataset. The RUL prediction results with uncertainty quantification at a 95 % confidence interval (CI) are also analyzed. The findings indicate that the proposed LSA + LSTM model, outperforms other optimization-based LSTM models in predictive accuracy, attaining a minimum Root Mean Square Error (RMSE) of 0.402 %, 0.526 %, 0.263 %, and 0.309 % for B₅, B₆, B₇, and B₁₈ batteries, respectively. In summary, this study offers a promising opportunity for RUL prediction of LIBs with uncertainty, thereby contributing to the harmless and effective operation of battery storage systems.