A noise-resilient adaptive deep learning framework for accurate state-of-charge prediction in lithium-ion batteries for electric vehicles

Abstract

Accurate estimation of the State-of-Charge (SoC) in lithium-ion batteries (LIBs) is essential for optimizing performance, ensuring safety, and prolonging battery life in Battery Management Systems (BMS) for Electric Vehicles (EVs). While Long Short-Term Memory (LSTM) networks have shown significant promise for SoC estimation, they often rely on manual hyperparameter tuning, leading to inconsistent accuracy and reduced adaptability. To overcome these limitations, this study introduces a robust, noise-resilient, and adaptive deep learning framework—MRFOSA-LSTM, that combines Manta Ray Foraging Optimization (MRFO) with Simulated Annealing (SA) to automate LSTM hyperparameter tuning. The hybrid MRFOSA enhances convergence and avoids local optima, while the addition of controlled noise during training improves the model’s robustness to external interference. The proposed method is rigorously analyzed and validated using multiple real-world driving cycles and evaluated across a wide range of initial SoC levels. Comparative analysis against baseline methods, including EKF, Particle swarm optimization (PSO) based LSTM, Genetic algorithm (GA) based LSTM, MRFO-LSTM, Transformer and Bi-LSTM methods, confirms the superior performance of MRFOSA-LSTM, achieving a Mean Absolute Error (MAE) of 0.25% and Root Mean Square Error (RMSE) of 0.36%. This framework offers a highly accurate and resilient solution for real-time SoC estimation in LIBs.