基于LSTM神经网络的锂离子电池荷电状态估算

针对锂离子电池荷电状态(state of charge, SOC)预测问题,采用长短期记忆循环神经网络(long short-term memory, LSTM)搭建电池SOC预测模型。利用直流电子负载对18650锂离子电池进行多工况放电,将电池电压、放电电流作为模型输入。将采集数据分为训练集、验证集和测试集,在训练集上训练模型,在验证集上调节模型超参数,在测试集上测试模型性能。采用带动量的随机梯度下降(stochastic gradient descent, SGD)进行权重更新,并加入Dropout正则化方法。在动态放电情况下,使用所提方法预测电池SOC最大绝对误差为2.0%,平均绝对误差为1.05%,验证了该方法的可行性。测试结果表明,在模型训练过程中加入Dropout正则化方法,可以有效降低网络的过拟合现象,增强模型的泛化能力。

Estimation on State of Charge of Lithium Battery Based on LSTM Neural Network

Aiming at a problem of prediction on the state of charge(SOC) of lithium-ion batteries, this paper makes use of the long short term memory(LSTM) neural network to build the SOC prediction model. By using DC electronic load for discharging under various working conditions for the 18650 lithium battery, it takes the battery voltage and discharge current as model inputs. It divides the collected data into the training set for training the model, the verification set for adjusting parameters of the model and the test set for testing model performance. It employs the stochastic gradient descent(SGD) with momentum for weight updating and adds the Dropout regularization method. The maximum absolute error of SOC by using the proposed method for prediction is 2.0% and the average absolute error(MAE) is 1.05% under dynamic discharge conditions, which has verified feasibility of the method. The results show that the Dropout regularization method can effectively reduce the over fitting phenomenon and enhance the generalization ability of the model.