基于实车数据和BP-AdaBoost算法的电动汽车动力电池健康状态估计

动力电池健康状态(state of health, SOH)估计是电动汽车领域关注的一个热点，目前的大部分方法都是基于实验室测试数据进行估计，忽略了实际车辆运行情况。使用国家大数据联盟平台采集的实际车辆运行数据进行电池SOH的估计。数据预处理方面，在清洗异常数据时，保留了实车数据中合理的强噪声数据，保证了数据的真实性。特征选择方面，选择容量增量曲线峰值和对应的电压以及基于安时积分得到的小片段充电容量数据。算法方面，针对真实数据的弱时序性问题，利用反向传播-自适应推进(back propagation-adapt boost, BP-AdaBoost)算法进行电池SOH估计的研究。最后，利用同一类型三辆车的数据进行了模型训练、测试和验证，预测结果与长短期记忆-循环神经网络(long short term memory-recurrent neural network, LSTM-RNN)算法对比，BP-AdaBoost算法估计误差更小，平均绝对误差达到0.96%,因此，所提出的方法可以应用于实车电池SOH的高精度估计。

Electric vehicle power battery SOH estimation based on real vehicle data and BP-Adaboost algorithm

State of Health (SOH) estimation is a hot topic in the field of electric vehicles. Most of the current methods are based on test data in the laboratory, so the actual vehicle operations are ignored. In this paper, the real vehicle operation data from the National Big Data Alliance platform was used to estimate SOH. In terms of data preprocessing, reasonable strong noise data in real vehicle data are retained to ensure the authenticity of data when rinsing abnormal data. In terms of feature selection, the peak value of capacity increment curve and corresponding voltage are selected as well as the small segment charging capacity data obtained based on ampere-hour integration. In terms of algorithm, BP-Adaboost algorithm is used to estimate SOH of battery for the weak timing of real data. Finally, the model is trained, tested and verified by using the data of three vehicles of the same type. Compared with LSTM-RNN algorithm, the estimation error of BP-Adaboost algorithm is smaller, and MAE can reach 0.96%. Therefore, the proposed method can be applied to high-precision SOH estimation of real vehicle batteries.