Hybrid deep learning framework for accurate state-of-charge estimation of electric vehicle batteries under multi-temperature conditions using transfer learning and WaOA optimization

Abstract

Accurate state-of-charge (SoC) estimation is essential for improving the safety, reliability, and energy management performance of lithium-ion batteries in electric vehicles, particularly under dynamic operating conditions and temperature variations. To address the limitations of existing data-driven methods in capturing nonlinear battery behavior and maintaining generalization across different thermal environments, this paper proposes a novel hybrid deep learning framework that integrates convolutional neural networks (CNN), bidirectional long short-term memory networks (BiLSTM), and an attention mechanism (AM), with model hyperparameters adaptively optimized using the Walrus Optimization Algorithm (WaOA). The novelty of the proposed approach lies in three aspects: first, the CNN–BiLSTM–AM architecture jointly exploits local feature extraction, bidirectional temporal dependency learning, and adaptive feature weighting for more informative SoC representation; second, WaOA is introduced to optimize key hyperparameters, improving convergence behavior and estimation robustness; third, a multi-temperature learning strategy combined with transfer learning is employed to enhance model adaptability under unseen temperature conditions. The proposed framework is evaluated using datasets collected at 0 °C, 25 °C, and 45 °C under standard driving cycles including DST, FUDS, and US06. Experimental results show that the proposed model achieves an average R² of 0.988, RMSE of 0.0203, MAPE of 3.98%, and MAE of 0.0163, while reducing RMSE by up to 12.1% and 49.2% compared with AOA-based and GWO-based benchmark models, respectively. In addition, the transfer-learning-based evaluation demonstrates strong cross-temperature generalization when the model is trained on combined-temperature data and fine-tuned on unseen 45 °C conditions. These findings confirm that the proposed WaOA-CNN-BiLSTM-AM framework provides an effective and robust solution for SoC estimation in temperature-varying electric vehicle battery systems. • A model structure for data-driven state of charge estimation is proposed. • The AM and CNN-BiLSTM modules are combined in parallel and optimized by WaOA • Model validated under fixed and varying temperatures across trained and unseen sets • Combined model enhances adaptability to unseen data by fine-tuning fully connected layers. • Results highlight the practicality of the model with enhanced estimation accuracy.