Lithium-ion Battery SOC Estimation based on CNN-Transformer Model

Abstract

The state of charge (SOC) of lithium-ion batteries is a critical state parameter in battery management systems (BMS), being closely associated with energy management, charge–discharge control and operational safety, yet it cannot be directly measured by conventional sensors. In recent years, deep learning methods have demonstrated considerable potential for SOC estimation owing to their strong capacity for nonlinear feature extraction. However, local dynamic characteristics and long-range temporal dependencies embedded in battery operating data are difficult to be modelled simultaneously by a single network architecture, thereby limiting SOC estimation accuracy under complex dynamic operating conditions. To address this limitation, a SOC estimation model integrating a multi-scale convolutional neural network (CNN) with a Transformer encoder is proposed. Local dynamic features from voltage and current sequences are extracted using multi-scale CNNs, while long-range temporal dependencies are further established by the Transformer encoder, thereby enhancing the representation of dynamic battery operating characteristics. During data preprocessing, SOC reference values are calculated using the ampere-hour integration method, and temporal input samples are constructed through normalization and a sliding-window strategy, providing a reliable basis for model training and prediction. Experiments are conducted using the publicly available CALCE INR 18650-20R lithium-ion battery dataset under four dynamic current profiles and three ambient temperatures. Model performance is evaluated using the root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R²). The proposed CNN-Transformer model achieves high SOC estimation accuracy, with an RMSE of 0.9371%, an MAE of 0.7913% and an R² of 0.9957 under the DST profile at 25 °C. Compared with CNN, bidirectional long short-term memory (BiLSTM) and Transformer models, the RMSE is reduced by up to 52.57%. Further evaluation across different temperatures and dynamic operating conditions demonstrates the favourable environmental adaptability and dynamic-condition robustness of the proposed model. These results establish an effective data-driven framework for high-precision SOC estimation and state monitoring in lithium-ion battery management systems.