基于改进图卷积网络的配电网状态估计方法

采用小样本量测信息进行配电网实时状态估计，对提高配电网可靠性，保证其稳定运行具有重要作用。为了在配电网量测信息不足条件下进行高精度状态估计，提出了一种基于改进的图卷积网络(graph convolutional network，GCN)的物理–数据融合配电网状态估计新方法。该方法首先利用少量相量量测单元将配电网进行切割分区，然后根据分区后的最大直径确定卷积网络所需要的卷积模块数量，其次修改了传统GCN中的邻接矩阵表示方法，从而实现利用卷积网络将配电网分区子系统中的状态变量均由量测量表示。通过IEEE33节点典型算例，验证了所提方法有效性。同时，通过与传统的高斯–牛顿优化算法和传统深度学习网络对比测试，结果表明，所提方法不仅能够将计算复杂度转移到离线阶段，而且能够不依赖于高冗余度的伪量测，具有较高的估计精度与计算速度。

An Improved Graph Convolutional Network-based State Estimation Method of Distribution Network

In order to improve reliability of distribution network and ensure its stable operation, the realtime state estimation of distribution network plays an important role in improving the reliability of distribution network and ensuring its stable operation. To perform state estimation with high accuracy under insufficient measuring information of distribution network, a new state estimation method for distribution network based on improved physical data fusion of graph convolutional network (abbr. GCN) was proposed. Firstly, in the proposed method a small number of phasor measuring units were utilized to divide the distribution network into partitions, then according to the maximum diameter after the partitioning the number of convolution modules required by the convolution network was determined. Secondly, the expressing method of traditional adjacent matrix in GCN was modified, thus by use of the convolution network all the state variables in the partitioned subsystems of distribution network were expressed by the measured quantities. By means of typical computing example of IEEE 33 bus system, the effectiveness of the proposed method was verified. Meanwhile, through the comparison test with traditional Gauss-Newton optimization algorithm and that with traditional deep learning network, the results show that using the proposed method not only the computation complexity can be transferred to off-line phase, but also it can be independent of the pseudo measurement with high redundancy, so the proposed method possesses higher estimation accuracy and computing speed.