Accurate prediction of the battery state of charge (SOC) is of great significance to improve the utilization efficiency and safety performance of the battery, and the monitoring of the battery state of charge is very important to help prevent overcharge and overdischarge accidents. The traditional SOC prediction methods are highly dependent on the mechanism model and statistical model, and have problems such as sensitive outliers and limited practical accuracy. In this study, a CNN-LSTM-AM (convolutional neural network - long short term memory neural network - attention mechanism) model is proposed to predict SOC variation trend through battery measurable variables.

The model first uses a one-dimensional convolutional neural network to extract spatial features of measurable variables, including battery current, voltage, temperature and average voltage, and then sends them to bidirectional long and short time memory for time series analysis. Finally, the attention mechanism is introduced to screen key features, reduce the redundancy of feature data, and improve the accuracy and generalization of the model. In addition, CNN-LSTM-AM model adopts rime optimization algorithm to optimize the hyperparameters in the training process, which effectively improves the training efficiency and reduces the training cost.

The actual evaluation on CALCE (Center for Advanced Life Cycle Engineering) data set of lithium iron phosphate shows that the attention mechanism can effectively improve the training performance of the prediction model, and the rime optimization algorithm adopted can help reduce the model hyperparameters, so as to obtain higher prediction accuracy. The performance of CNN-LSTM-AM model was tested under different temperature conditions, and both RMSE and MAE were less than 1%, which was sufficient to confirm the feasibility of the model to predict SOC. In addition, even if the initial SOC is uncertain, the proposed CNN-LSTM-AM model can still accurately track SOC trend changes, and the overall prediction accuracy reaches RMSE<1.5% and MAE<1.5%. The RMSE and MAE results of the network proposed in this study are smaller than those of CNN-LSTM and CNN-LSTM-AM. It shows strong robustness and generalization ability. Finally, in order to comprehensively compare the performance of different SOC prediction methods, the CNN-LSTM-AM model proposed in this study is compared with other experimental results. It can be seen that the method proposed in this study has significantly lower RMSE compared with AT-CNN-LSTM. At the same time, considering that the proposed method uses less training set data, we can also see the advantages of the designed network. Compared with EI-LSTM-CO(extended input-LSTM-constrained output), it can be found that the error is close. In addition, EI-LSTM-CO performs some post-processing on the predicted SOC, which can also reflect the superiority of the proposed method.

The following conclusions are drawn from the simulation analysis: (1) A CNN-LSTM-AM model is proposed and applied to the SOC prediction task of battery, which can effectively capture important input features and improve the prediction accuracy. (2) Design a rime optimization algorithm, which can automatically search the optimal solution of CNN-LSTM-AM model, effectively reduce the time cost of hyperparameter optimization. (3) The influence of different ambient temperatures and initial SOC values on the prediction accuracy of CNN-LSTM-AM was studied, and the performance of CNN-LSTM-AM was compared with that of traditional prediction models to verify its strong robustness and high generalization ability.