Abstract:

The driving mileage prediction of pure electric vehicles is one of the most concerns for drivers. The existing regression models always have the drawbacks of low prediction accuracy and large relative error. This paper developed a machine learning method that combines segment regression prediction and single- point classification prediction to predict the mileage. The prediction method took real vehicle state parameters, environmental information as input, extracted the optimal feature set by clustering and filtering encapsulated feature selection, then selected a prediction method based on the sample size of driving segments, and layered coupling prediction of environmental temperature and battery health state (SOH) to improve the prediction accuracy of fragment regression. The final prediction result was further optimized by the model fusion of single point classification prediction and fragment regression prediction. The RMSRE relative error of the predicted result of the mileage test set is 0.035, and the average relative error is 1.71% , which can accurately and stably achieve the mileage prediction.

Key words: urban traffic, prediction of driving mileage, data driving, electric vehicle, model fusion

摘要：

纯电动汽车行驶里程预测是驾驶者最关心的问题之一，为解决现有预测算法模型精度低、相对误差大的问题，本文采用融合片段回归与单点分类的机器学习方法对行驶里程进行预测.以真实车辆各项状态参数、环境信息等作为输入，通过聚类和过滤封装式特征筛选，提取最优特征集合，并基于行驶片段样本量选择预测方法，通过对环境温度和电池健康状态(SOH)进行分层耦合提高片段回归预测精度，通过单点分类和片段回归预测模型融合优化最终预测结果.行驶里程测试集预测结果中均方根相对误差(RMSRE)为0.035，平均相对误差为1.71%，能够精确稳定地实现行驶里程预测.

关键词: 城市交通, 行驶里程预测, 数据驱动, 电动汽车, 模型融合