	US 12,235,311 B1
	Graph theory-based method and device for locating instability fault source in direct-current microgrid
Xin Zhang, Hangzhou (CN); Xueqi Liu, Hangzhou (CN); Fanfan Lin, Hangzhou (CN); Hao Ma, Hangzhou (CN); and Yuankui Wang, Hangzhou (CN)
Assigned to ZHEJIANG UNIVERSITY, Hangzhou (CN)
Filed by ZHEJIANG UNIVERSITY, Hangzhou (CN)
Filed on Jun. 25, 2024, as Appl. No. 18/753,853.
Claims priority of application No. 202410272786.6 (CN), filed on Mar. 11, 2024.
Int. Cl. G01R 31/08 (2020.01)

CPC G01R 31/088 (2013.01) [G01R 31/086 (2013.01)]

6 Claims

1. A graph theory-based method for locating an instability fault source in a direct-current microgrid, comprising the following steps:

measuring and obtaining electrical data of each power electronic converter in the direct-current microgrid, and processing the electrical data to obtain a corresponding feature vector;

constructing a graph model based on the direct-current microgrid, wherein the graph model takes the power electronic converters as nodes, power flow paths among all the power electronic converters as edges among the nodes, and the feature vectors corresponding to the power electronic converters as a node feature matrix; wherein a specific structure of the graph model is as follows:

setting G=(V, E, X) to represent an attribute graph in the graph model, wherein

V represents a set of the nodes, that is, one power electronic converter represents one node; and N is a total number of the nodes;

E represents a set of the edges among the nodes, that is, the power flow paths among all the power electronic converters are taken as node edges; an adjacency matrix A represents a set of the edges among the nodes and meets A ∈ custom character

^N×N; when there is a power flow between a node i and a node j and i≠j, a matrix A_ijis set to 1, and otherwise the matrix A_ijis set to 0; and in addition, A_iiis set to 1; and

X represents the node feature matrix, and meets X ∈ custom character

^N×C; and C represents feature dimensions of the nodes;

according to whether the direct-current microgrid is unstable, labeling the electrical data of the power electronic converters, and combining the electrical data and labels of the power electronic converters into data sets;

introducing the graph model into a pre-constructed graph convolutional neural network framework to construct a corresponding classification network, which comprises a feature extraction module, a feature fusion module, and a classification module, wherein

the feature extraction module is configured to obtain electrical change data of the power electronic converters and position information in the corresponding graph model to generate a feature vector corresponding to the electrical change data and a node vector corresponding to the position information;

the feature fusion module is configured to generate a corresponding node feature representation according to the generated feature vector and node vector; and

the classification module is configured to output a classification result according to the input node feature representation, wherein the classification result comprises whether a direct-current grid is unstable and an instability fault source; wherein an expression of the node feature representation is as follows:

wherein D represents a degree matrix, each

A represents an updated adjacency matrix, A ∈ custom character

^N×N; X represents the node feature matrix, X ∈ custom character

^N×C; N represents a total number of the nodes; W represents a weight matrix, W ∈ custom character

^C×F; F represents feature vector dimensions extracted by all the convolution layers; and H represents node feature representation extracted by convolution of all the convolution layers, H ∈ custom character

^N×F;

wherein the weight matrix is constructed by comparing feature vectors of two nodes connected by each edge between nodes, which is specifically as follows:

if feature vectors of the nodes at both ends of the edge are different, defining a weight as 1; and

if feature vectors of the nodes at both ends of the edge are same, defining the weight as 0;

utilizing the data set to train the classification network to obtain a grid diagnosis model for instability diagnosis of the direct-current grid; and during training, parameters of the classification network are updated by using a cross entropy loss function, wherein an expression of the cross entropy loss function is as follows:

wherein ŷ_irepresents a predicted label of a node i, y_irepresents an actual label of the node i, and N is a total number of the nodes;

inputting the electrical data of each power electronic converter of the direct-current grid to be diagnosed into the grid diagnosis model;

outputting a judgment result of whether the direct-current grid is unstable and an existing instability fault source;

locating one or more faulty power electronic converters;

replacing the one or more faulty power electronic converters;

wherein the power electronic converters include direct-current to direct-current converters for adjusting voltage levels, inverters for converting direct-current to alternating-current, and bidirectional converters for controlling energy storage; and

wherein types of instability faults are converter single operation instability or converter cascade operation instability.