A data-driven framework for lithium-ion battery RUL using LSTM and XGBoost with feature selection via Binary Firefly Algorithm

Time-series analysis algorithms have been developed to directly forecast the Remaining Useful Life (RUL) of Lithium-ion batteries, as opposed to indirect methods that operate through the estimation of the State of Health (SOH) based upon extracted health indicators (HIs). This indirect approach facilitates a more detailed exploration of the connection between battery health indicators and RUL decay, potentially enhancing the precision of predictions compared to methods that solely model the RUL. In this study, 16 Lithium-ion battery HIs are extracted and analyzed by the Long Short-Term Memory (LSTM) model. The forecasting outcomes of the LSTM models are selected by the heuristic optimization algorithm Binary Firefly Algorithm (BFA) to formalize the feature subset that yields the highest prediction accuracy for the gradient boosting algorithm eXtreme Gradient Boosting (XGBoost) to estimate future SOH values as RUL. The efficacy of the proposed methodology is validated using the Lithium-ion battery dataset from NASA's Prognostics Center of Excellence, where the experiments show a 23 % improvement in RMSE and MAPE compared to direct RUL modeling using the same time-series analysis algorithm. Furthermore, the model built on data from one battery was successfully applied to other batteries under similar operating conditions, achieving comparable forecasting accuracy, and the qualitative ramifications of these findings are discussed.