State of Charge Estimation for Power Battery Packs Under Multiple Temperature Conditions Based on a Hybrid Neural Network Model

Accurate estimation of the State of Charge (SOC) in power battery packs is one of the core technologies in electric vehicle energy management systems. To address the nonlinear characteristics of power batteries under dynamic operating conditions and the error accumulation issues of traditional methods, this paper proposes a hybrid model (CNN-BiLSTM-Attention-KF) that integrates a Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory Network (BiLSTM), Attention Mechanism, and Kalman Filter. The model extracts local features from signals such as voltage and current using CNN, captures long-term dependencies in time series through Bi-LSTM, and enhances the weight of key features with the Attention Mechanism. Subsequently, the Kalman Filter is introduced to dynamically optimize the prediction results, suppressing noise interference and error accumulation. Experimental results under multi-temperature environments and dynamic stress test conditions demonstrate that the proposed method reduces the root mean square error of SOC estimation by 38.6 % compared to the traditional Bi-LSTM model, with the mean absolute error consistently remaining below 0.6 %. This significantly improves estimation accuracy and robustness under multi-temperature operating conditions.