College of Electronic Engineering, Guangxi Normal University
Guangxi Science and Technology Major Project (AA20302003)
为了提高锂电池荷电状态SOC(state of charge)的预测精度，提出一种基于注意力机制和CNN-LSTM融合模型的锂电池荷电状态预测方法。该模型采用一维卷积神经网络CNN(convolution neural network)和长短时记忆LSTM(long short-term memory)神经网络学习得到SOC与锂电池放电数据的非线性关系，以及SOC序列存在的长期依赖性。同时，该模型采用“多对一”的结构，将当前时刻的锂电池SOC和多个历史时刻的放电数据建立映射关系，并通过注意力机制关注到对当前时刻的SOC影响较大的历史放电数据，进一步提升SOC预测的准确度。动态工况下的锂电池SOC预测实验表明，该方法在不同温度条件下的平均预测误差为0.89%，与SVM、GRU、XGBoost相比，分别降低了81.2%、66.7%和56.5%，且优于未融合注意力机制的LSTM和CNN-LSTM，具有较高的预测精度和应用价值。
To improve lithium battery state of charge (SOC) prediction accuracy, a method based on the fusion model of attention mechanism and CNN-LSTM was proposed. The model uses one-dimensional convolution neural network (CNN) and long short-term memory (LSTM) neural network to learn the nonlinear relationship between SOC and lithium battery discharge data, and the long-term dependence among SOC? sequences. At the same time, the model adopts a "many-to-one" structure to establish the mapping relationship between current SOC and discharge data at multiple historical moments, and pays attention to the historical discharge data which has greater influence on the current SOC through the attention mechanism to further improve the SOC prediction accuracy. The SOC prediction experiments under dynamic conditions show the average prediction error of the method is 0.89% under different temperature conditions, which is 81.2%, 66.7% and 56.5% lower than that of SVM, GRU and XGBoost respectively, and the method is also superior to ablation model such as LSTM and CNN-LSTM and has higher prediction accuracy and application value.