Generalizable and Efficient State of Charge Estimation for Lithium-Ion Batteries with Deep Learning

Accurate state of charge (SOC) estimation is crucial for reliable operation of lithium-ion batteries. Present SOC estimation techniques are poorly generalized for batteries of different chemical properties and inefficient in model training. In this paper, we construct a pre-trained model that combines a convolutional neural network (CNN) with a self-attentive mechanism to realize more efficient and generalized SOC estimation. CNN and self-attention mechanisms are utilized to extract more general features from battery operation. And a model is pre-trained, followed by transfer learning, to achieve accurate SOC estimation with reduced training costs. Experimental results show that the proposed pre-trained model achieves 0.97% mean absolute error (MAE) and 1.02% root mean square error (RMSE) using only three minutes of fine-tuning training. In addition, the proposed method outperforms other deep learning methods in terms of accuracy and efficiency. We demonstrate that the proposed method is feasible under different conditions, including different material, temperature, and operation conditions.