基于高斯过程回归的中压配电网雷击过电压预测方法

Prediction Method of Lightning Overvoltage in Medium VoltageDistribution Network Based on Gaussian Process Regression

准确评估雷击过电压对合理绝缘配合研究至关重要,基于此,提出了一种基于机器学习的高斯过程回归模型,用于预测大范围地电阻率和介电常数下的雷击过电压。即采用二维时域有限差分技术,结合地面场线耦合模型,计算了雷击过电压。考虑到输入空间的可变性,考虑了7个预测变量,即电阻率、介电常数、回击速度、闪点距离、回击峰值电流、前沿时间和架空线路的高度,以确定响应变量,即雷击过电压。因此,高斯过程回归模型的训练是使用从二维时域有限差分技术获得的雷击过电压对第一次和随后的雷击进行的。利用均方根误差对模型的估计精度进行了评估,表明指数核函数在描述协方差方面的误差最小。最后,通过将预测模型与二维时域有限差分技术生成的随机测试案例进行对比,评估预测性能,预测结果与时域有限差分法的预测结果吻合较好,验证了所提方法的有效性。

Accurate evaluation of lightning overvoltage is very important for the study of reasonable insula⁃tion coordination.A gauss process regression model based on machine learning is proposed to predict lightning overvoltage under a wide range of ground resistivity and dielectric constant.The lightning over⁃voltage is calculated by two-dimensional finite-difference time-domain technique combined with the ground-field line coupling model.Considering the variability of the input space,seven predictive varia⁃bles,namely resistivity,dielectric constant,return speed,flashpoint distance,return peak current,leading edge time and height of overhead line,were considered to determine the response variable,name⁃ly,lightning overvoltage.Thus,the Gauss process regression model was trained for the first and subse⁃quent lightning strikes using the lightning overvoltage obtained from the two-dimensional finite-diffe-rence time-domain technique.The estimation accuracy of the model is evaluated by root mean square error,which shows that the error of exponential kernel function is the smallest in describing covariance.Finally,the prediction performance is evaluated by comparing the prediction model with the random test case generated by the 2D FDTD technique.The prediction results are in good agreement with the predic⁃tion results of the FDTD method,which verifies the effectiveness of the proposed method.