State of charge estimation for lithium-ion battery using a multi-feature Mamba network and UKF under mixed operating conditions

With the global focus on sustainable energy, lithium-ion batteries have emerged as a key energy storage solution. Accurate state of charge (SOC) estimation is crucial for battery performance, lifespan, and safety, yet remains challenging under mixed operating conditions. This paper proposes a hybrid method combining a multi-feature encoding fusion Mamba (MFMamba) neural network with an unscented Kalman filter (UKF) for robust and accurate SOC estimation. The MFMamba network designs a multi-feature extraction module to capture latent information from current, voltage, and temperature sequences, utilizing a selective mechanism and attention mechanism to learn feature information under varying conditions. The UKF is then integrated to filter noise from the predictions. Experiments on public datasets with three batteries under single and mixed operating conditions at various temperatures demonstrate the superiority of the proposed method over mainstream machine learning approaches. Under mixed operating conditions, the maximum MAE and RMSE are 0.92% and 1.32%, respectively, while under single operating condition, they are 0.67% and 0.75%, respectively. The proposed method effectively learns battery sequence feature information under different conditions, enabling accurate and robust SOC estimation.