Active distribution network fault location method based on S-TGCN

Abstract

The present invention discloses an active distribution network fault location method based on S-TGCN, which introduces a spatiotemporal attention mechanism and a spatiotemporal convolutional layer to extract the global spatiotemporal features of the distribution network fault period from the time series of the graph structure, thereby improving the model's global fault feature extraction capability in scenarios with high-proportion distributed power access, and its fault location accuracy reaches 98.76%. Through the feature aggregation and differentiated pooling function of S-TGCN, the robustness of the new method in scenarios with missing data and data noise is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of distribution network fault location, and in particular relates to an active distribution network fault location method based on S-TGCN. Background Art

[0002] Short-circuit faults in distribution networks can easily lead to power outages, resulting in huge economic losses and social impacts. In order to improve the reliability of distribution network operation, the fault location of the distribution network should be located promptly and accurately to facilitate isolation of the fault area and rapid restoration of power supply. In recent years, wide-area measurement systems based on a large number of micro-phasor measurement units have effectively improved the observability of distribution systems, making it possible to locate distribution network faults based on machine learning.

[0003] Machine learning methods have good learning ability and can make accurate judgments on the fault location using the voltage and current collected by the measuring devices before and after the fault. Sapountzoglou N, Lago J, De Schutter B, et al proposed a fault location method for active distribution networks based on deep neural networks. The voltage and current of each node before and after the fault and the power generation and load in the distribution network were used as the input of the model to locate the fault section, and good results were achieved. Deng X, Yuan R, Xiao Z, et al used wavelet transform to extract eigenvalues ​​from zero-sequence current. Then, support vector machine (SVM) was used to identify the fault section in the distribution network. However, this method only analyzes single-phase grounding faults. Jin Q, Ju R proposed to divide the distribution system into main branches and independent areas. Genetic algorithms were used to find main branch faults and independent area faults. Existing machine learning methods have achieved certain research results in the location of distribution network faults, but the above methods use Euclidean space data as input to train the fault location model, without considering the real topological structure of the distribution network and ignoring the correlation characteristics between nodes.

[0004] With the emergence of graph neural network (GNN), neural network methods have been extended to the analysis of graph data in non-Euclidean space. The combination of deep learning and graph data has been preliminarily realized. Spatial-Temporal Graph Convolutional Network (S-TGCN) introduces temporal convolution on the basis of graph convolution, and has good classification performance in the scenario of dynamic changes in the attributes of each node. Chen K, Hu J, Zhang Y, et al proposed to use graph convolutional network (GCN) for fault location in distribution network. Since it considers the system topology and integrates multiple measurements on different buses, it has high fault location accuracy. In order to improve the adaptability of fault location model to topology changes, Li Jiawei, Wang Xiaojun, He Jinghan, Zhang Yongjie, Zhang Dahai proposed to use graph attention network (GAT) in the distribution network fault location method based on graph attention network to realize distribution network fault location. The fault location accuracy under the topology change of distribution network is improved by changing the adjacency matrix and dynamically updating the attention coefficient between adjacent vertices. However, the above method does not consider the impact of distributed generation on the fault location of the distribution network. After the high-power distributed generation is connected to the distribution network, the voltage and current values ​​of each node during the fault period are significantly affected

[12] , which reduces the separability between samples.

[0005] In addition, there are measurement and communication errors in actual working conditions, which lead to data noise and missing problems in the measurement data of each node in the distribution network during the fault period. Sapountzoglou N, Lago J, De Schutter B, et al proposed to improve the matrix algorithm by adding directional fault judgment criteria in A generalizable and sensor-independent deep learning method for fault detection and location in low-voltage distribution grids [J], thereby improving the computational efficiency and information fault tolerance of the matrix method. Patel B proposed to use fast discrete orthogonal S transform to extract fault current features in A new FDOSTentropy based intelligent digital relaying for detection, classification and localization of faults on the hybrid transmission line [J], and used SVM as a classifier to improve the fault tolerance of fault location. Mishra DP, Ray P. proposed to use multiple machine learning algorithms in Fault detection, location and classification of a transmission line [J] to reduce the influence of noise and outliers on fault location accuracy by aggregating results. However, the positioning fault tolerance of the above methods is easily affected by data noise and the location and type of nodes with missing data. Therefore, it is still necessary to study high-fault-tolerant fault location methods. Summary of the invention

[0006] The purpose of the present invention is to provide an active distribution network fault location method based on S-TGCN, which can enhance the robustness in data missing and data noise scenarios.

[0007] The technical solution adopted by the present invention is an active distribution network fault location method based on S-TGCN, which is specifically implemented according to the following steps:

[0008] Step 1: Establish an S-TGCN model, including an input layer, two spatiotemporal convolutional layers, a pooling layer, a parallel dual-channel differential pooling layer, a fully connected layer, and an output layer;

[0009] Step 2: Combine the distribution network topology information and the measurement data of each node before and after the fault to construct a time-space graph sequence data reflecting the real topology of the distribution network;

[0010] Step 3: Use the time-space attention mechanism and the time-space convolution layer to adaptively extract the global time-space features of the distribution network fault period;

[0011] Step 4: Use a parallel dual-channel differentiated pooling function to jointly extract and integrate multiple global features, and output the active distribution network fault location through the fully connected layer and the output layer.

[0012] The present invention is also characterized in that:

[0013] The measurement data of each node before and after the fault in step 2 specifically refers to the three-phase voltage, three-phase current, and active and reactive power injected into each node before and after the fault.

[0014] In step 1, the parallel dual-channel differential pooling layer includes a maximum pooling feature layer and a differential pooling feature layer, the maximum pooling feature layer is the first channel, and the differential pooling feature layer is the second channel.

[0015] The global feature vector of the maximum pooling feature layer is:

[0016] M{ρ 1 ,ρ 2 ,…,ρ N}=

[0017] max(F(h 1 ),F(h 2 ),…,F(h n )) (1)

[0018] In the formula, (F(h 1 ),F(h 2 ),…,F(h n )) is the middle-level feature matrix after the time-space convolution layer, and max is the maximum value function;

[0019] The feature vector of the differential pooling feature layer is expressed as:

[0020] A{ρ 1 ,ρ 2 ,…,ρ N}=

[0021] avg(F(h 1 ),F(h 2 ),…,F(h n )) (2)

[0022] Where avg is the mean function.

[0023] Step 2: The distribution network topology information refers to the connection relationship between nodes. If a node is connected to an adjacent node, it is marked as 1, otherwise it is marked as 0. The matrix formed by 1 and 0 to represent the connection relationship of nodes in the distribution network is used as the adjacency matrix.

[0024] The spatiotemporal convolution layer includes graph convolution and standard 2D convolution. Graph convolution extracts spatial features by aggregating the information of neighboring nodes in the distribution network nodes, while standard 2D convolution extracts temporal features by utilizing the temporal dependency of the features of each node between adjacent time frames.

[0025] Graph convolution is defined as:

[0026]

[0027] g θ Represents the convolution kernel. *G represents the graph convolution operation. represents a Chebyshev coefficient vector, x is the input data, L = DA represents the Laplace matrix, is the degree matrix, λ max is the maximum eigenvalue of the Laplace matrix, I N is the identity matrix, T k (x) = 2xT k-1 (x)-T k-2 (x) is a recursively defined Chebyshev polynomial, and T 0 (x) = 1, T 1 (x) = x.

[0028] The specific process of step 3 is as follows:

[0029] Step 3.1, extracting the global time-space fault features of the distribution network through the time-space convolution layer;

[0030] Step 3.2: In the spatial domain, the spatial attention mechanism participates in the operation of the graph convolution layer through the spatial attention matrix to obtain a graph structure representation of the fault line expression capability;

[0031] Step 3.3: In the time domain, the temporal attention mechanism is used to capture dynamic temporal dimension information.

[0032] The spatial attention mechanism in step 3.2 is defined as follows:

[0033] P=V p ·σ((χ (l-1) Z 1 )Z 2 (Z 3 χ (l-1) ) T +b p ) (4)

[0034] P′ m,n =softmax(P m,n ) (5)

[0035] in, represents the trainable parameters of the model, σ represents the sigmoid activation function, represents the input of the lth layer, P represents the temporal attention matrix, P m,n Represents the correlation between node m and node n;

[0036] The graph structure representation of the fault line expression capability is obtained according to the correlation between nodes.

[0037] The definition of the temporal attention mechanism in step 3.3 is as follows:

[0038] Q=V q ·σ(((χ (l-1) ) T M 1 )M 2 (M 3 χ (l-1) )+b q ) (6)

[0039] Q′ m,n =softmax(Q m,n ) (7)

[0040] in represents the trainable parameters, Q m,n Represents the correlation strength between time frame m and time frame n. The time attention matrix Q is normalized using the softmax function. The input of S-TGCN is regulated by the time attention mechanism.

[0041] The specific process in step 4 is:

[0042] The global feature extracted by the differential pooling function of each channel is expressed as:

[0043]

[0044] In the formula T is the merged high-level global feature, is the concatenation operator;

[0045] Multiple global features are extracted and synthesized together to obtain:

[0046] Where θ(N)=M{ρ 1 ,ρ 2 ,…,ρ N}, T{p 1 ,p 2 ,…,p 2n} is the global feature after integration;

[0047] The integrated global features are input into the fully connected layer, and finally the fault location of the active distribution network is output through the output layer.

[0048] The beneficial effects of the present invention are:

[0049] (1) The temporal-spatial attention mechanism and the temporal-spatial convolutional layer are introduced to extract the global temporal-spatial features of the distribution network fault period from the time series of the graph structure, which improves the model's ability to extract global fault features in scenarios with a high proportion of distributed power generation access, and its fault location accuracy reaches 98.76%.

[0050] (2) Through the feature aggregation and differentiated pooling function of S-TGCN, the robustness in data missing and data noise scenarios is enhanced. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is a schematic diagram of the overall framework structure of the S-TGCN model of the present invention;

[0052] Figure 2 It is a schematic diagram of input data of the present invention;

[0053] Figure 3 This is a wiring diagram of the distribution network containing distributed photovoltaic IEEE33 nodes in the case of the present invention.

[0054] Figure 4 This is a visualization comparison of hidden layer features of the four methods using t-SNE in the case of the present invention.

[0055] Figure 5 It is a diagram of the fault line location capability of different models in the case of the present invention for the scenario of the ratio of the total active output value of distributed photovoltaic to the total load active value.

[0056] Figure 6 This is the accuracy and one-hop accuracy diagram of the S-TGCN model in the case of the present invention under strong data noise and missing scenarios. DETAILED DESCRIPTION

[0057] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0058] The active distribution network fault location method based on S-TGCN of the present invention is implemented in the following steps:

[0059] Step 1: Establish the overall architecture of the S-TGCN model as follows Figure 1 As shown, it includes an input layer, two spatiotemporal convolutional layers, a pooling layer, a parallel dual-channel differential pooling layer, a fully connected layer, and an output layer;

[0060] Among them, the parallel dual-channel differential pooling layer includes a maximum pooling feature layer and a differential pooling feature layer, the maximum pooling feature layer is the first channel, and the differential pooling feature layer is the second channel;

[0061] In the fault localization model, the pooling layer performs pooling operations on the extracted fault localization temporal-spatial features and generates a global feature vector;

[0062] One of the points h j After convolution, the feature mapped to the middle layer is Γ N =F(h j ). In the formula, Γ N is the mapped feature vector, N is the feature vector length, and F is the feature mapping function. The global feature vector of the maximum pooling feature layer is:

[0063] M{ρ 1 ,ρ 2 ,…,ρ N}=

[0064] max(F(h 1 ),F(h 2 ),…,F(h n )) (1)

[0065] In the formula, (F(h 1 ),F(h 2 ),…,F(h n )) is the middle-level feature matrix after the time-space convolution layer, and max is the maximum value function;

[0066] The differential pooling feature layer takes the mean of a certain dimension of data as the feature value of that dimension. The feature vector of the differential pooling feature layer is expressed as:

[0067] A{ρ 1 ,ρ 2 ,…,ρ N}=

[0068] avg(F(h 1 ),F(h 2 ),…,F(h n )) (2)

[0069] Where avg is the mean function.

[0070] The spatiotemporal convolution layer includes graph convolution and standard 2D convolution. Graph convolution extracts spatial features by aggregating the information of neighboring nodes in the distribution network nodes, while standard 2D convolution extracts temporal features by utilizing the temporal dependency of the features of each node between adjacent time frames.

[0071] Graph convolution is defined as:

[0072]

[0073] g θ Represents the convolution kernel. *G represents the graph convolution operation. represents a Chebyshev coefficient vector, x is the input data, L = DA represents the Laplace matrix, is the degree matrix, λ max is the maximum eigenvalue of the Laplace matrix, I N is the identity matrix, T k (x) = 2xT k-1 (x)-T k-2 (x) is a recursively defined Chebyshev polynomial, and T 0 (x) = 1, T 1 (x) = x.

[0074] In active distribution network fault location, the functions of each part are:

[0075] The input layer inputs the time-space graph sequence data composed of the measurement data of each node during the fault period and the adjacency matrix. The color difference is used in the graph to reflect whether the node is connected to the distributed power supply. In the time-space convolution layer, the spatial attention and spectral method graph convolution layer are first used to aggregate the spatial correlation characteristics between adjacent nodes. Among them, the graph convolution based on the spectral method is to project the graph structure data into the spectral domain space through Fourier transform and then perform convolution operation, which avoids the difficulty of convolution definition caused by the graph data not satisfying translation invariance. It is suitable for the perception and analysis of global features; because the topological structure of the distribution network is relatively sparse and the in-degree of each node is low, the graph convolution operation method based on the spectral method is used to aggregate the features of adjacent nodes; then, the features of each node are extracted in the time dimension through the time attention mechanism and the standard 2D convolution to extract the time-varying features of the distribution network nodes. In the pooling layer, in order to ensure the robustness of the model in data noise and missing scenarios, a parallel dual-channel pooling layer is used to extract differential global features to achieve high-tolerance and accurate fault location. The first channel extracts global features by the maximum pooling function, and the other channel uses a differential pooling function, and its global features are extracted by the maximum pooling and mean pooling functions. The pooled global features are flattened into a one-dimensional vector by the Flatten layer and passed to the fully connected layer. The output of the fully connected layer is obtained through the softmax operation to obtain the fault line.

[0076] Step 2: Combine the distribution network topology information and the measurement data of each node before and after the fault to construct a time-space graph sequence data reflecting the real topology of the distribution network; wherein the measurement data of each node before and after the fault specifically refers to the three-phase voltage, three-phase current, and active and reactive power injected by each node before and after the fault;

[0077] The distribution network topology information refers to the connection relationship between nodes. If a node is connected to an adjacent node, it is marked as 1, otherwise it is marked as 0. The matrix formed by 1 and 0 to represent the connection relationship of nodes in the distribution network is used as the adjacency matrix.

[0078] The time-space graph sequence data reflecting the real topology of the distribution network is input to the input layer:

[0079] Define an undirected graph G = {ν, ε, A}. ν is the set of distribution network nodes, |v| = N, N is the number of distribution network nodes. ε is the edge set. A is the adjacency matrix, which reflects the connection relationship between nodes in the distribution network. Each image frame contains the adjacency matrix A and the node feature X that reflects the fault line information of the distribution network.

[0080] The sampling frequency of each node feature is 100Hz. The node feature is composed of the three-phase voltage time series V a ,V b ,V c , three-phase current time series I a ,I b ,I c , the active time series P of each node power output G and reactive time series Q G composition.

[0081] Where L represents the number of features of each node, T S Indicates the number of image frames for each sample. T S =30, that is, all node feature data in the period from 0.1s before the fault to 0.2s after the fault occurs are taken to form the fault sample. There are two main reasons for this design. One is to ensure that the fault is in the early stage of its steady state, and the other is to assume that no protection element is activated within 0.2s after the fault occurs, so as to isolate the fault. i ∈X(i∈{1,2,…,T S}) represents the features of each node at time i.

[0082]

[0083] Among them I a,k (k∈{1,2,…,N}) represents the effective value of phase A current at node k, I b,k represents the effective value of the B-phase current at node k, I c,k represents the effective value of the C-phase current at the k-node, U a,k represents the effective value of the A phase voltage at node k, U b,k represents the effective value of the B-phase voltage at node k, U c,k represents the effective value of the C-phase voltage at node k, P G,k Indicates the active output value of the k-node power supply, QG,k Represents the reactive power output value of the power supply at node k. If the node has no power supply, then P G,k and Q G,k is 0.

[0084] In order to let the model learn the real topological structure of the distribution network, the adjacency matrix A is constructed as the input of the model. When the topology of the distribution network changes, the adjacency matrix also changes accordingly to adapt to the change of topology.

[0085]

[0086] Where i,j∈{1,2,…,N}. Figure 2 The figure shows the input of the S-TGCN model when the network topology changes. The output y is the classification label of the sample, indicating the fault line or normal operation state corresponding to the sample.

[0087] Step 3: Use the time-space attention mechanism and the time-space convolution layer to adaptively extract the global time-space features of the distribution network fault period; the specific process is as follows:

[0088] Step 3.1: Extract the global spatiotemporal fault features of the distribution network through the spatiotemporal convolution layer; the spatiotemporal graph convolution layer consists of graph convolution and standard 2D convolution. Graph convolution extracts spatial features by aggregating the information of neighboring nodes in the distribution network nodes, and 2D convolution extracts temporal features by using the temporal dependency of the features of each node between adjacent time frames.

[0089] The spectrum-based graph convolution operation is used to extract the spatial features of each node in the distribution network in the spatial dimension, and the adjacency matrix A is constructed according to the actual topological structure of the distribution network. The Chebyshev expansion of the graph Laplacian operator is used to reduce the computational complexity. The definition of the K-1 order Chebyshev graph convolution is:

[0090]

[0091] g θ Represents the convolution kernel. *G represents the graph convolution operation. represents a Chebyshev coefficient vector. x is the input data. L=DA represents the Laplace matrix. is the degree matrix. λ max is the largest eigenvalue of the Laplace matrix. N is the identity matrix. k (x) = 2xT k-1 (x)-T k-2 (x) is a recursively defined Chebyshev polynomial, and T 0 (x) = 1, T 1 (x) = x.

[0092] The approximate expansion of Chebyshev polynomials is used to extract the information of 0 to K-1 neighbors centered on each node.

[0093] The above definition is extended to nodes with multiple features. The input of the lth layer is C l-1 Indicates the number of channels of each node. l-1 represents the time dimension of the lth layer. For each X i Both use C l The convolution kernel performs convolution operation on it to get g θ *GX i . is the convolution kernel parameter. Therefore, the information of 0 to K-1 connected nodes is aggregated to each node.

[0094] In order to obtain the time-varying characteristics of each node information during the fault period, after the graph convolution operation fully extracts the spatial features of each moment, a standard 2D convolution layer is used to extract the time-varying characteristics of the node information. The time domain convolution operation on the lth layer is defined as:

[0095]

[0096] Where ReLU is the activation function, Φ represents the parameters of the convolution kernel, and * represents the standard convolution operation.

[0097] Step 3.2: In the spatial domain, the spatial attention mechanism participates in the operation of the graph convolution layer through the spatial attention matrix to obtain a graph structure representation of the fault line expression capability;

[0098] In the spatial domain, since the fault location is random, different line segment faults have different effects on the fault characteristics of each node in the distribution network. The spatial attention mechanism participates in the operation of the graph convolution layer through the spatial attention matrix to obtain a graph structure representation with better fault line expression ability. The spatial attention mechanism is defined as follows:

[0099] P=V p ·σ((χ (l-1) Z 1 )Z 2 (Z 3 χ (l-1) ) T +b p ) (5)

[0100] P′ m,n =softmax(P m,n ) (6)

[0101] in, represents the trainable parameters of the model, σ represents the sigmoid activation function, represents the input of the lth layer, P represents the temporal attention matrix, P m,n Represents the correlation between node m and node n.

[0102] When performing graph convolution, the model dynamically updates the node weights by combining the adjacency matrix A and the spatial attention matrix P that reflect the actual topological structure of the distribution network.

[0103] The graph structure representation of the fault line expression capability is obtained according to the correlation between nodes.

[0104] Step 3.3: In the time domain, there is a correlation between the node feature changes between adjacent time frames, and the correlation will change under different circumstances. Therefore, the temporal attention mechanism is used to capture dynamic temporal information;

[0105] The temporal attention mechanism is defined as follows:

[0106] Q=V q ·σ(((χ (l-1) ) T M 1 )M 2 (M 3 χ (l-1) )+b q ) (7)

[0107] Q′ m,n =softmax(Q m,n ) (8)

[0108] in represents the trainable parameters, Q m,n Represents the correlation strength between time frame m and time frame n. The time attention matrix Q is normalized using the softmax function. The input of S-TGCN is regulated by the time attention mechanism. Pay more attention to valuable time dimension information.

[0109] Step 4: Use the parallel dual-channel differentiated pooling function to jointly extract multiple global features and synthesize them, and output the active distribution network fault location through the fully connected layer and the output layer; the specific process is:

[0110] The maximum pooling feature layer and the differential pooling feature layer can obtain different significant features of the measurement data of each node. A parallel dual-channel pooling layer is used to extract global features. The first channel uses the maximum pooling function to extract global features, and the other channel uses the differential pooling function. Its global features are extracted by the maximum pooling and differential pooling functions to improve the robustness of the model. The global features extracted by the differential pooling function of each channel are expressed as:

[0111]

[0112] In the formula T is the merged high-level global feature, is the concatenation operator;

[0113] Multiple global features are extracted and synthesized together to obtain:

[0114] Where θ(N)=M{ρ 1 ,ρ 2 ,…,ρ N}, T{p 1 ,p 2 ,…,p 2n} is the global feature after integration;

[0115] The integrated global features are input into the fully connected layer, and finally the fault location of the active distribution network is output through the output layer.

[0116] Example

[0117] The PSCAD / EMTDC platform is used to build an IEEE33 node distribution network system connected to distributed photovoltaics. The overall topology of the distribution network is as follows: Figure 3 As shown, the voltage level is 12.66 kV, the frequency is 60 Hz, the total load active power is 3.725 MW, and the reactive power is 2.3 MVar. It is worth noting that the connecting line segments in the topology here only represent the connection relationship between the buses, and their length is not the geometric connection distance between the buses.

[0118] The system has 33 nodes, 32 loads, 5 distributed photovoltaics, 32 distribution lines and 5 tie lines. The power factor of distributed photovoltaics is 1. Its specific access location and capacity are shown in Table 1. Such configuration of distributed photovoltaics can reduce the power loss of the distribution system and improve its voltage distribution.

[0119] Table 1

[0120] PV Number 1 2 3 4 5 Access location bus number 5 8 15 17 27 Access capacity (kW) 1430 233 275 370 335

[0121] Distributed photovoltaics are composed of photovoltaic arrays, photovoltaic power generation maximum power tracking control modules, inverters, and LC filter circuits. The inverter obtains the reference voltage through the instantaneous reactive power theory, and uses the space vector pulse width modulation (SVPWM) method to obtain the pulses required to control the switch, thereby achieving the purpose of controlling the inverter.

[0122] A large number of input data samples of the described fault period model are generated through simulation to train the fault location model and verify its robustness under different working conditions. The simulation case configuration is shown in Table 2. The fault simulation of the distribution network considers the following different impacts.

[0123] (1) Fault line: Since the location of short-circuit faults is random in the actual operation of the distribution network, a distribution line in the IEEE33-node distribution network is simulated as the fault line.

[0124] (2) Fault location: The distance between the fault point and the head node of the fault line is considered (the node with the larger node number among adjacent nodes is defined as the head node, for example, the head node of the distribution line between nodes (1 and 18) is node 18). The ratio of the impedance between the fault point and the head node of the fault line to the impedance of the fault line is used to reflect the distance ratio between the fault point and the head node of the fault line.

[0125] (3) Fault type: 10 fault types in 5 categories are considered, including single-phase grounding, two-phase short circuit grounding, two-phase short circuit, three-phase short circuit grounding, and three-phase short circuit.

[0126] (4) Fault impedance: In order to study the influence of fault impedance on fault location, eight fault impedance values ​​are set as shown in Table 2.

[0127] (5) Distributed photovoltaic power output state: The photovoltaic installed capacity is used as the standard value. Due to the large volatility of distributed photovoltaic output, six distributed photovoltaic output states are considered, and each photovoltaic output fluctuates randomly within the range of ±15% of the total output state value.

[0128] (6) Load level: The load fluctuates randomly within the range of ±15% of the total load state value, but the total load output state remains at the specified level.

[0129] Therefore, the simulation generates a total of 138,240 fault samples (the fault sample parameter values ​​are shown in Table 2). In addition, only 100 normal samples are simulated and generated under different distributed photovoltaic and different load levels, and a total of 1,800 normal operation samples are generated.

[0130] The total data set contains 140,040 faulty and normal samples, which are used in model training, validation, and testing.

[0131] Table 2

[0132]

[0133]

[0134] The specific implementation steps of the S-TGCN model training process are as follows: 70% of the samples are used for training, 10% of the samples are used for validation, and 20% of the samples are used for testing. At the same time, the 10-fold cross-validation method is used to train the model. The optimal model is selected as the distribution network fault location model. Softmax Loss is used as the loss function to update the trainable parameters of the model. This model is built based on the Tensorflow framework, the hardware platform is Windows 10, the CPU is AMD Ryzen9 5900X 12-CoreProcessor, the GPU is Nvidia GeForce RTX 3060, and the RAM size is 32G.

[0135] The size of the temporal and spatial convolution kernels of this model is 10. The batch size is 66. In order to avoid overfitting, a dropout layer is added with a dropout rate of 50%. The initial learning rate is 0.001, and the adam optimizer is used to optimize it in real time during the training process. The number of cycles is 500.

[0136] Three baseline models were implemented and compared while ensuring that the training and test data were consistent:

[0137] (1) Support vector machine (SVM). The measured data is reduced to 100 dimensions through principal component analysis (PCA) and input into the support vector machine to determine the fault location. The Gaussian kernel function is used as the kernel function of the SVM.

[0138] (2) Convolutional neural network (CNN). CNN has 4 convolutional layers, 2 fully connected layers and 1 output layer. The feature extraction uses the maximum pooling function.

[0139] (3) Graph convolutional network (GCN). GCN has 3 graph convolution layers, 2 fully connected layers and 1 output layer.

[0140] Fault location is actually a graph classification problem. The accuracy is used as an evaluation index to evaluate the fault location model. Define tp as the number of samples where the faulty line segment is correctly identified, and tn as the number of samples where the healthy running samples are correctly identified. fp is the number of samples where the faulty line segment is incorrectly identified, and fn is the number of samples where the healthy running samples are incorrectly identified. The accuracy is defined as:

[0141]

[0142] In addition to the accuracy rate, the one-hop accuracy rate is also used as an indicator to measure the performance of the model, that is, if the predicted faulty line is directly connected to the actual faulty bus, the faulty line is considered to be located correctly.

[0143] Table 3

[0144]

[0145] Table 3 shows the accuracy and one-hop accuracy of various methods in the test set. The S-TGCN model has the highest positioning accuracy, which verifies the effectiveness of the proposed method. The accuracy of the PCA+SVM and CNN methods is relatively low, but their one-hop accuracy is still satisfactory.

[0146] Since faults occur randomly in the scenario of differentiated load demand and distributed power output, the voltage and current of each node during the fault period have a high level of uncertainty, which reduces the separability between fault samples and between fault samples and normal operation samples. This puts higher requirements on the fault location model to capture the characteristics of the fault line and improve its classification ability.

[0147] In order to evaluate the impact of distributed power output on fault location, the ratio of distributed photovoltaic total active output to total load active output is defined as P rate .

[0148]

[0149] Where P PV is the total active power output of distributed photovoltaics, P Load is the total load active value. According to the fault sample simulation configuration in Table 2, the test set data is divided into rate The numerical intervals are divided into 4 groups: [0,20%), [20%,40%), [40%,60%), [60,80%).

[0150] In order to evaluate the model’s ability to extract features from input data, t-distributed stochastic neighbor embedding (t-SNE) is used to reduce the dimensionality of high-dimensional data and visualize it in two-dimensional space. To evaluate the impact of distributed generation access on fault location, the following Figure 3 The adjacent line fault samples of the distributed power supply access location are shown, with a total of 10 fault lines. rateUnder the premise that the number of numerical interval samples is the same, 2500 samples are randomly selected from each fault line in the test sample for display. In order to verify the feature extraction capabilities of different methods, the PCA data, the fault data after the 4th convolution layer of CNN, the fault data after the 3rd graph convolution layer of GCN, and the fault data after the 2nd time-space graph convolution layer of S-TGCN are extracted and reduced to two-dimensional space for display. The results are shown in Figure 4 As shown in Figure 2, the tighter the distribution of samples of the same fault line, the better the effect of the model in extracting features.

[0151] According to the experimental results, for short-circuit fault data in the scenario where distributed power sources are connected to the distribution network, S-TGCN has better fault feature extraction capability and stronger separability between samples, which improves the fault location capability of the model.

[0152] Figure 5 The fault line location capabilities of different models are demonstrated for scenarios where the ratio of total distributed photovoltaic active output to total load active output varies. Figure 5 , in different P rate In the scenario of numerical interval, S-TGCN has the highest fault location accuracy. The fault location accuracy of the GCN model decreases rapidly as the proportion of distributed power output increases. The CNN and PCA+SVM methods have relatively low fault location accuracy.

[0153] The distribution system measurement device has problems such as measurement and communication errors in actual working conditions, which leads to data noise and potential data loss in each node data. In order to verify the effectiveness of the proposed method in actual working conditions, the simulation data of the test set is modified as follows:

[0154] (1) Adding Gaussian Noise (GN): Add Gaussian noise with different signal-to-noise ratios (SNR). The mean of the noise is 0 and the standard deviation is σ.

[0155] (2) Node Data Missing (NDM): N nodes are randomly removed from each sample. drop The data of each node is replaced with 0. drop =[1,2,3,4,5].

[0156] (3) Set random data missing (RDM): All measurement data are lost with probability P. Missing data are supplemented with 0. P = [0.01, 0.02, 0.03, 0.04, 0.05].

[0157] In order to evaluate the impact of data noise and missing types on fault location of various methods, P = 0.01 was used to modify the test set data.

[0158] Table 4 shows the fault line location accuracy and one-hop accuracy of the S-TGCN model and other baseline models under the above noise and data loss levels. As shown in Table 4, the S-TGCN model has the highest accuracy compared with other methods in various data noise and data loss scenarios. Among them, node data loss has the greatest impact on the fault accuracy and one-hop accuracy of each model. Since GCN and S-TGCN aggregate the feature information between adjacent nodes, their one-hop accuracy can reach more than 90% in the node loss scenario, but the positioning accuracy of PCA+SVM and CNN methods is lower.

[0159] Table 4

[0160]

[0161] To evaluate the robustness of the S-TGCN model under strong data noise and missing scenarios. According to σ, N drop , P values ​​are different, and the test set data is modified to create 5 different scenarios:

[0162] Scenario 1:

[0163] Scenario 2:

[0164] Scenario 3:

[0165] Scenario 4:

[0166] Scenario 5:

[0167] Figure 6 The accuracy and one-hop accuracy of the S-TGCN model in scenarios with strong data noise and missing data are demonstrated. According to the experimental results, as the degree of data noise and missing data increases, the fault location accuracy of S-TGCN decreases. However, due to the use of the time-space convolution layer and the time-space attention mechanism to extract the global features of the fault period and perform differential pooling feature extraction, in scenario 5 where the data noise and missing data are the most serious, the one-hop accuracy of the new model can still reach 90.18%.

[0168] Through the above method, the present invention introduces the time-space attention mechanism and the time-space convolution layer to extract the global time-space characteristics of the distribution network fault period from the time series of the graph structure, and improves the global fault feature extraction ability of the model in the scenario of high-proportion distributed power access, and its fault location accuracy reaches 98.76%. Through the feature aggregation and differentiated pooling function of S-TGCN, the robustness of the new method in the scenario of data missing and data noise is enhanced.

Claims

1. The active distribution network fault location method based on S-TGCN is characterized by: Follow the steps below to implement it: Step 1: Establish an S-TGCN model, including an input layer, two spatiotemporal convolutional layers, a pooling layer, a parallel dual-channel differential pooling layer, a fully connected layer, and an output layer; The parallel dual-channel differential pooling layer in step 1 includes a maximum pooling feature layer and a differential pooling feature layer, the maximum pooling feature layer is the first channel, and the differential pooling feature layer is the second channel. The global feature vector of the maximum pooling feature layer is: （1） In the formula, is the middle-level feature matrix after the time-space convolution layer, is the maximum value function; The feature vector of the differential pooling feature layer is expressed as: （2） Where avg is the mean function; Step 2: Combine the distribution network topology information and the measurement data of each node before and after the fault to construct a time-space graph sequence data reflecting the real topology of the distribution network; The measurement data of each node before and after the fault in step 2 specifically refers to the three-phase voltage, three-phase current, and active and reactive power injected into each node before and after the fault; Step 3: Use the time-space attention mechanism and the time-space convolution layer to adaptively extract the global time-space features of the distribution network fault period; Step 4: Use a parallel dual-channel differentiated pooling function to jointly extract and synthesize multiple global features, and output the active distribution network fault location through the fully connected layer and the output layer; The specific process in step 4 is: The global feature extracted by the differential pooling function of each channel is expressed as: （8） In the formula , is the merged high-level global feature, is the concatenation operator; Multiple global features are extracted jointly and synthesized to obtain: , where , is the global feature after integration; The integrated global features are input into the fully connected layer, and finally the fault location of the active distribution network is output through the output layer.

2. The active distribution network fault location method based on S-TGCN according to claim 1 is characterized in that: The distribution network topology information in step 2 refers to the connection relationship between nodes. If a node is connected to an adjacent node, it is marked as 1, otherwise it is marked as 0. The matrix formed by 1 and 0 to represent the connection relationship of nodes in the distribution network is used as the adjacency matrix.

3. The active distribution network fault location method based on S-TGCN according to claim 1 is characterized in that: The time-space convolution layer includes graph convolution and standard 2D convolution. The graph convolution extracts spatial features by aggregating information of neighbor nodes in the distribution network node, and the standard 2D convolution extracts temporal features by utilizing the temporal dependency of the features of each node between adjacent time frames.

4. The active distribution network fault location method based on S-TGCN according to claim 3 is characterized in that: The graph convolution is defined as: （3） represents the convolution kernel, represents the graph convolution operation, represents a Chebyshev coefficient vector, is the input data, L=DA represents the Laplace matrix, is the degree matrix, , is the largest eigenvalue of the Laplacian matrix, is the identity matrix, are recursively defined Chebyshev polynomials, and , .

5. The active distribution network fault location method based on S-TGCN according to claim 1 is characterized in that: The specific process of step 3 is as follows: Step 3.1, extracting the global time-space fault features of the distribution network through the time-space convolution layer; Step 3.2: In the spatial domain, the spatial attention mechanism participates in the operation of the graph convolution layer through the spatial attention matrix to obtain a graph structure representation of the fault line expression capability; Step 3.3: In the time domain, the temporal attention mechanism is used to capture dynamic temporal dimension information.

6. The active distribution network fault location method based on S-TGCN according to claim 5 is characterized in that: The spatial attention mechanism in step 3.2 is defined as follows: （4） （5） in, , , , represents the trainable parameters of the model, represents the sigmoid activation function, Indicates The input of the layer, represents the temporal attention matrix, Representation Node and nodes The correlation between The graph structure representation of the fault line expression capability is obtained according to the correlation between nodes.

7. The active distribution network fault location method based on S-TGCN according to claim 5 is characterized in that: The definition of the temporal attention mechanism described in step 3.3 is as follows: （6） （7） in , , , represents the trainable parameters, express Timeframe and The correlation strength of the time frame, using the softmax function to normalize the time attention matrix , the input of S-TGCN is regulated by the temporal attention mechanism .

Images