基于1D-CNN的含DG中压馈线合环电流预测方法

Loop Closing Current Prediction Method of Medium Voltage Feeder with DG Based on 1D-CNN

摘要: “双碳”目标下分布式电源(distributed generator, DG)大量接入配电网，使得传统基于物理建模的合环电流计算方法难以满足工程对计算准确性、适用性的要求。为充分挖掘合环电流输入特征的时空联系，解决输入特征对合环电流的不稳定影响导致的预测精度下降问题，提出了一种基于一维卷积(1D-CNN)的含DG配电网合环电流预测方法。首先，结合实际配电网运行特点分析了合环电流输入特征构建，提出了两种合环电流预测的输入特征选择；其次，基于DIgSILENT/PowerFactory搭建仿真模型形成样本集合；最后，分别对两种输入特征进行模型训练，并对超参数寻优及预测流程等问题进行了分析，从而建立1D-CNN合环电流预测模型。在贵州某地区实际中压配电系统开展仿真分析，结果显示该模型在馈线a、b首端及合环处的电流测试集样本Ia、Ib、Ic上的平均绝对误差分别为0.0927%、0.2704%和0.4797%，表明所提方法能准确且稳定预测合环电流。

Abstract: A large number of distributed generators (DG) are connected to the distribution network under the “dual-carbon” goal, which makes the traditional loop closing current calculation method based on physical modeling difficult to meet the requirements of engineering for calculation accuracy and applicability. In order to fully explore the spatio-temporal relation of input features of the loop closing current and solve the problem of prediction accuracy decline caused by the unstable influ-ence of input features on the loop closing current, a method for predicting the loop closing current of distribution network with DG based on one-dimensional convolutional neural network (1D-CNN) is proposed. Firstly, the input characteristics of the loop closing current are analyzed in combination with the operational characteristics of the actual distribution network, and two kinds of input char-acteristics for loop closing current prediction are proposed; then, a simulation model is constructed based on DIgSILENT/PowerFactory to form the sample set; and finally, the model training of the two input features is performed separately, and hyperparameter optimization and prediction process are analyzed, so as to establish the 1D-CNN loop closing current prediction model. The simulation analysis then is carried out in the actual medium voltage distribution system in a region of Guizhou. The results show that the mean absolute errors of the model on the current samples Ia, Ib and Ic are 0.0927%, 0.2704% and 0.4797% respectively, indicating that the proposed method is capable of correctly and stably predicting the loop closing current.