A Model Based on Convolutional Long Short-Term Memory and Attention Mechanism for Predicting the State of Charge of Batteries

This paper introduces a deep learning model that integrates convolutional neural netswork (CNN) and Long Short-Term Memory (LSTM) networks, enhanced by an attention mechanism, for accurate State of Charge (SOC) estimation in lithium-ion batteries. The CNN component captures local spatial features such as periodic variations in current and voltage, while the LSTM network models the temporal dependencies inherent in battery data. The attention mechanism prioritizes significant time points, further improving the model's ability to focus on critical temporal features influencing SOC. Experimental results demonstrate that the CNN&LSTM&Attention model significantly outperforms the CNN&LSTM and LSTM models across various operating conditions and temperature ranges. For instance, under Beijing Dynamic Stress Test (BJDST) conditions, the CNN&LSTM&Attention model achieves an MSE (Mean Squared Error) of 0.0002, compared to 0.0003 for CNN&LSTM and 0.0006 for LSTM, with corresponding RMSE (Root Mean Squared Error) and MAE (Mean Absolute Error) values of 0.0130 and 0.0101, respectively, which are the lowest among the models. Similarly, under FUDS (Federal Urban Driving Schedule) and US06 (United States Driving Cycle 06) cycles, the hybrid model consistently exhibits lower MSE, RMSE, and MAE values. At 0°C, it achieves an MSE of 0.0002 and a maximum error of 0.0695, outperforming CNN&LSTM (MSE: 0.0004, Max Error: 0.0729) and LSTM (MSE: 0.0009, Max Error: 0.1700). Even at 45°C, where battery behavior becomes more challenging to model, the hybrid approach delivers superior predictions with a maximum error of 0.1817, compared to 0.2233 for CNN&LSTM and 0.2210 for LSTM. These results highlight the superior prediction accuracy and generalization capability of the CNN&LSTM&Attention model, making it a robust and reliable solution for SOC estimation under diverse conditions.