基于遗传算法改进后的反向传播神经网络的实车电池全时刻健康度预测

Real-Vehicle Battery All-Time State of Health Prediction Based on Genetic Algorithm-Optimized Backpropagation Neural Network

摘要:

基于纯电动汽车实车数据，提出一种融合安时积分法与基于遗传算法改进的反向传播（GA-BP）神经网络的动力电池全时刻健康度（SOH）预测方法。首先采用安时积分法计算深度充电片段电池容量值，处理离群点、低频数据和温度因素造成的误差，经曲线拟合得到全部片段电池容量值，再通过平均滑动滤波和插值法计算全部数据采集时刻电池SOH。然后根据不同预测时刻前的某次充电是否符合深度充电片段特征分别训练不同的GA-BP神经网络模型。最终形成包含电池容量计算、充电片段判别及模型预测在内的电池全时刻SOH预测方法。基于3辆车的实车数据进行验证，结果表明该方法可使用实车数据对纯电动汽车全时刻电池SOH进行预测，保证精度的同时训练数据量需求较小。

Abstract:

A genetic algorithm-backpropagation（GA-BP） neural network integrated with the ampere-hour integration method is proposed for all-time state of health （SOH） prediction of power batteries using actual electric vehicle data. First， the ampere-hour integration method calculates battery capacity during deep charging segments， while addressing outliers， low-frequency data artifacts， and temperature-induced errors， through curve fitting to obtain complete segment capacity values. Sliding average filtering and interpolation are then applied to derive SOH values at all data acquisition points. Then， distinct GA-BP neural network models are trained based on whether preceding charging segments before prediction timestamps exhibit deep charging characteristics. Finally， an integrated framework encompassing capacity calculation， charging segment discrimination， and model prediction is established for comprehensive SOH estimation. Validation with real-world data from three vehicles demonstrates that the method accurately predicts SOH at all times while requiring minimal training data.