基于多元线性回归的锂动力电池荷电状态鲁棒预测

Robust Prediction of state-of-charge Battery Based on Multiple Linear Regression

摘要:

目前广泛使用的锂电池荷电状态(state-of-charge, SOC)预测方法的训练数据需要通过大量的仿真实验获取，而电动汽车在充电过程中产生的大量的充电记录数据并没有得到合理利用。为了能有效利用这些充电记录数据，将多元线性回归算法应用到SOC预测中。多元线性回归方法将电压、电流、电容等物理量作为与SOC直接相关的输入变量从而对SOC进行回归预测。由于SOC的时序特征，将SOC预测分为多个子预测过程，不断迭代计算，循环预测SOC的下一时刻输出值。同时为了克服异常样本对SOC预测精度的影响，采用两种常见的鲁棒回归算法(Theil-sen算法与RANSAC算法)来进行SOC预测。实验结果表明，鲁棒回归算法及多元线性回归算法能够很好地捕捉到SOC的增长规律，相比之下，Theil-sen算法精度更高，误差约1.398%，能够很好地满足SOC预测的实际需求。

Abstract:

The training data of the widely adopted lithium battery state-of-charge (SOC) prediction method needs to be obtained through plenty of simulation experiments, while the large amount of charging record data generated by the electric vehicle during the charging process is not properly utilized. In order to effectively utilize these data, a multiple linear regression algorithm is applied to the SOC prediction. The multivariate robust regression method uses the physical quantities such as voltage, current, and capacitance of the lithium battery which directly related to the SOC as input variables to perform regression prediction on the SOC. Due to the timing characteristics of the SOC, the prediction of SOC is divided into multiple sub-prediction processes, and the iterative calculation is continuously performed to cyclically predict the next output value of SOC. At the same time, two common robust regression algorithms (Theil-sen algorithm and RANSAC algorithm) are used to predict SOC in order to overcome the influence of abnormal samples on SOC prediction accuracy. The experimental results show that the robust regression algorithm and the multiple linear regression algorithm can well capture the growth law of SOC. In contrast, Theil-sen algorithm has higher precision and the error is about 1.398%, which can satisfy the actual needs of SOC prediction well.