A closed-loop state of charge estimation of lithium-ion batteries by integrating deep learning and adaptive Kalman filter

Under various operating environments and temperature conditions, accurately estimating the state of charge (SOC) of batteries is a crucial prerequisite for an efficient battery management system. However, the SOC of lithium-ion batteries exhibits complex nonlinear and dynamic characteristics influenced by multiple external factors. Furthermore, the measurement noise and system uncertainties encountered in real-world applications will degrade the accuracy of SOC estimation. Therefore, this study proposes a closed-loop SOC estimation method that integrates deep learning with filters. The method first constructs a hybrid deep learning model to obtain a noisy SOC observation. Specifically, a temporal convolutional network (TCN) is used to extract time-domain features from the raw data, and a frequency-domain channel attention mechanism (FECAM) is embedded into the TCN architecture to capture key features across frequency-domain channel dimensions. Subsequently, to further improve estimation accuracy and outlier resistance, a closed-loop SOC estimation is achieved using a noise-adaptive Kalman filter (AKF) on the noisy SOC observation and prediction. Finally, the experimental results demonstrate that the proposed TCN-FECAM-AKF method achieves higher estimation accuracy under different temperatures and the Federal Urban Driving Schedule driving cycle, with a root mean square error and maximum error of less than 2.40% and 5.93%, respectively.