Enhanced group convolutional hybrid neural network for state of charge estimation of lithium-ion batteries consider temperature uncertainty

Accurate state of charge (SOC) estimation for lithium-ion batteries is a pivotal area of research within battery management systems. Despite the high predictive accuracy achievable through machine learning techniques, significant output volatility remains a challenge. To address these challenges, this study introduces a novel high-precision SOC estimation methodology that leverages machine learning. Specifically, a data augmentation method using local weighted regression algorithm is proposed. The augmented data is used for neural network training, and this technique effectively mitigates output fluctuations. Furthermore, a group convolutional neural network model integrated with multilayer self-attention mechanisms, optimized via a proportional-integral-derivative search algorithm for hyperparameter tuning. The data enhanced hybrid neural network model demonstrates superior prediction accuracy and reduced output variability. Ablation and comparative experimental results validate the proposed method, achieving an average MAE of less than 0.65 %, an average RMSE of less than 0.76 %, and an average R2 exceeding 0.9991 under BBDST and DST working conditions in wide temperature range. Under NEDC working condition at dynamic temperature, MAE was 0.676 %, RMSE was 1.011 % and R2 was 0.98633. This innovative approach provides an outstanding and stable SOC estimation solution for lithium-ion batteries.