Remaining Useful Life Prediction of EV Lithium-Ion Batteries Using TG-AMF Enhanced MTS-BiLSTM Optimized with Walrus Algorithm

Abstract

Predictive maintenance is essential to industrial operations, especially in lithium-ion battery systems. Remaining Useful Lifetime (RUL) prediction is the most accurate means of maintaining optimal performance and avoiding unexpected failure in these systems. Noisy sensor data and complex degradation patterns, however, make the task complex and require sophisticated techniques for proper analysis and forecasting. Hence, this manuscript proposes an optimized Electric Vehicle (EV) lithium-ion battery RUL prediction using a multi-time scale bidirectional Long Short-Term Memory (MTS-BiLSTM) Network. Initially, input data is taken from the Remaining Useful Lifetime Prediction Dataset available on Kaggle, containing sensor readings and operational parameters for systems under degradation. It uses the trimmed global adaptive mean filter approach (TG-AMF), which enhances input data quality by removing general noise and preserving the essential feature points related to degradation patterns. A novel MTS-BiLSTM network approach is proposed for predicting RUL by capturing both Proximal correlations and Delayed dependencies in data to model trends accurately. Hyperparameters of the proposed model are optimized via the Walrus Optimizer (WO), which improves prediction and computation overhead. The robustness of the proposed framework is analyzed through benchmarking with performance metrics coefficient of determination (R2), root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), prediction horizon accuracy (PHA), and computation time (CT). The overall PHA of 98%, MAE of 17.04, MAPE of 8.63, CT of 7.57s, and RMSE of 5.83 are obtained by the proposed method of forecasting the battery RUL for EVs.