Fault Location Method for Distribution Networks Based on SimAM-GraphSAGE-GAT

Abstract

In distribution networks, traditional fault location methods have insufficient anti-interference capability and low accuracy in locating high-resistance grounding faults. To address these issues, a distribution network fault location method on the basis of SimAM-GraphSAGE-GAT is proposed. Firstly, the distribution network topology structure is converted into an adjacency matrix, and the electrical parameters of the faulty line are incorporated as node features into the graph structure of the network. Subsequently, the sampling and aggregation mechanism of GraphSAGE is used for learning node representation. Features are refined using SimAM. As a parameter-free attention mechanism, SimAM improves the ability of the model to capture important fault information. Then, the multi-head attention mechanism of GAT is introduced to enhance the representation of neighborhood relationships. Finally, GraphSAGE is utilized once again for deep aggregation, with a view to localizing faults by node classification. An IEEE 33-node distribution network model is adopted to verify the effectiveness of the algorithm in the experiment. The results show that this method can maintain high positioning accuracy even under the tested conditions, such as high-resistance grounding, noise interference, and data loss.

1. Introduction

With the development of the Chinese economy, the power system is becoming increasingly large and complex. Modern energy systems are rapidly evolving toward AI-driven, digital, and integrated operations. As the intermediate link connecting the main power grid to end users, the safe and stable operation of distribution grids ensures the reliable functioning of the power system [1,2,3]. With the continued growth of distributed generation (DG) such as photovoltaic (PV) systems, its large-scale integration into the grid can alter power-flow directions, transforming distribution networks from traditional single-source radial networks into multi-source networks with bidirectional power flows. As a result, the operating environment of distribution networks becomes more complex, placing higher demands on their safe and stable operation, as well as on operation and maintenance. However, under increasingly complex operating conditions, ensuring the safe operation of the underlying distribution network during faults is essential for supporting the advanced data-driven operation of modern energy systems. Rapid and accurate fault location in power systems is of great significance for shortening power outage durations, reducing economic losses, and improving service quality [4,5,6].

Currently, distribution networks in China mostly employ a low-current grounding system with non-directly grounded neutral points [7,8]. In low-current grounding systems, single-phase ground faults take up approximately 80% of all faults. If not addressed immediately, such faults may escalate into more serious phase-to-phase short-circuit faults, which pose significant risks to the safe and stable operation of the power system [9]. In traditional distribution network fault location methods, Zeng et al. [10] used injected non-power-frequency characteristic current signals to establish amplitude comparison criteria between faulted and non-faulted sections and thus achieved fault location. However, its reliability is susceptible to high transition resistance. Wang et al. [11] applied the continuous wavelet transform to aerial-mode traveling wave analysis and identified faulty branches by comparing the similarity between the fault-time–frequency matrix and a preset template. Then, they utilized the energy evolution spectrum of a specific frequency band combined with the double-ended method for fault location. Nevertheless, this application is challenging in distribution networks due to the strict requirements for double-ended synchronous measurements. Chandran et al. [12] used the synchronous voltage and current phasors provided by the synchronous phasor measurement unit (PMU) for narrowing the fault area using apparent impedance. Next, multi-terminal measurement was employed to construct an extended impedance equation set to solve for the accurate fault location. In high-resistance faults, however, the location error shows a significant increase and may even fail to provide the correct fault location.

To tackle the above issues, artificial intelligence technologies have been applied to fault location [13,14]. Mirshekali et al. [15] performed spectro-frequency analysis on voltage signals to generate two-dimensional features and achieved fault location via a deep learning framework combining convolutional neural networks with capsule networks. Zhou et al. [16] leveraged cloud theory to enhance the genetic algorithm and optimize the structure of the backpropagation (BP) neural network. In addition, fault state variables were used as the input to achieve fault location in distribution networks. While the methods described in the above literature are effective under certain conditions, their adaptability is limited in complex operating conditions. These methods struggle to make use of the critical network structural characteristics inherent in distribution networks [17].

For the past few years, researchers have begun to approach the issue of fault location in distribution networks from a new perspective with the rapid advancement of graph neural networks [18]. Chen et al. [19] integrated graph convolutional neural networks with node electrical properties and topological features and enhanced adaptability in complex scenarios. Wang et al. [20] utilized the graph sample and aggregate (GraphSAGE) algorithm to extract power system structure and electrical information and established a mapping relationship between node features and fault labels. This approach reduces computational complexity to diagnose and locate faults. Liang et al. [21] enhanced the expressive power of the model for neighborhoods by calculating weights between adjacent nodes through a multi-head attention mechanism. The methods employed in the above-mentioned literature all focus on the aggregation of information between nodes, with limited capacity to distinguish the importance of different features within a node. Yang et al. [22] proposed using a simple attention module (SimAM) for enhancing key features, but SimAM is predominantly applied in convolutional neural networks. Systematic integration within graph neural networks and studies on its application in distribution network fault location scenarios are still relatively scarce [23]. The summary of the relevant research is presented in

Table 1. Summary of relevant works.

Method	Main Limitation	The Solution Proposed in This Paper
Signal-injection/traveling-wave/impedance-based methods	Sensitive to fault resistance, synchronization requirements, or measurement conditions	The proposed method learns fault-related patterns directly from node electrical features and topology information
Conventional AI-based methods	Hard to fully capture the structural correlation between nodes in distribution networks	This paper accomplishes the modeling of topological relationships by representing the distribution network as a graph structure
GraphSAGE-based methods	Important electrical features may be weakened during neighborhood aggregation	SimAM is introduced to recalibrate feature importance before deeper graph aggregation
GAT-based methods	Limited ability to distinguish the importance of different electrical feature dimensions within a node	SimAM enhances fault-sensitive dimensions

.

To resolve the above issues, a fault location method for distribution networks is proposed in this paper based on SimAM-GraphSAGE-graph attention network (GAT). Although the increasing grid-connection penetration of distributed generators (DGs) has rendered the operating conditions of distribution networks more dynamic and complex, the underlying physical topology of a standard distribution network still constitutes the fundamental framework for analyzing the state correlation between nodes and fault propagation. On this basis, the validation of the proposed method is mainly carried out for the operating conditions of the benchmark distribution network studied in this paper. The topology of the distribution network is mapped into a graph structure, and three-phase voltage, three-phase current, active power, and reactive power are fed into the model as node features. GraphSAGE is adopted to construct initial node representations with local topological constraints, and SimAM is used to perform parameter-free recalibration on fault-sensitive feature dimensions. The multi-head attention mechanism in GAT is applied to implement differentiated weighted aggregation of information from different neighborhoods so as to compensate for the critical information that may be lost during the aggregation process. Finally, GraphSAGE is used to further integrate high-order local topological information and complete node fault discrimination.

2. Principles of the SimAM–GraphSAGE–GAT Model

2.1. Structured Representation of Distribution Network Diagrams

The distribution network graph model can be expressed as G = (V,E), where V and E represent nodes and edges, respectively. The adjacency matrix A and the node feature matrix X stand for inputs for a fault localization model. The adjacency matrix A denotes an N × N symmetric matrix representing the graph topological structure calculated in the following way:

$$A(i,j)=\left\{\begin{array}{c}1,i\ne j,(v_i,v_j)\in E\\ 0,else\end{array}\right.$$

(1)

The node feature matrix X represents an N × F matrix, where N stands for the total number of nodes and F indicates the number of features per node. Node features contain active power, reactive power, three-phase voltage, and three-phase current. They are expressed as

$$(V_1,V_2,V_3,I_1,I_2,I_3,P,Q)$$

(2)

The feature vector for each node is represented as

$$X=\{V_1,V_2,V_3,I_1,I_2,I_3,P,Q\},X\in R^{N\times 8}$$

(3)

2.2. GraphSAGE Neighbor Sampling and Preliminary Feature Extraction

When a fault occurs in the distribution network, fault information propagates to adjacent nodes along the lines connected to the faulty node. To characterize the changes in node state correlation caused by single-line-to-ground faults within the local topological range, this layer adopts GraphSAGE to aggregate the electrical information of the node itself and that of its first-order neighborhood, so as to construct initial node feature representations under local topological constraints. The core of the GraphSAGE algorithm lies in its inductive neighbor-sampling and feature-aggregation mechanism. It randomly samples a set number of neighbor nodes, fuses the feature information of the node itself with that of its first-order neighbor nodes, and thus learns the local topological information of the nodes. The learning process of the algorithm is presented in Figure 1.

For node i in the graph, its layer L feature is $h_i^{(L)}$ In the first layer, $h_i^{(0)}$ represents the raw features of the node when L = 0, which correspond to a total of eight-dimensional electrical physical quantities of node i, including three-phase voltage, three-phase current, active power, and reactive power. The neighbor features of node i are aggregated using mean aggregation. The aggregated neighbor features can be expressed as

$$m_i^{(0)}={\displaystyle \frac{1}{\left|N(i)\right|}}\sum _{u\in N(i)}{h}_{u_j}^{(0)}$$

(4)

where N(i) denotes the set of neighbors of node i. During the stage of feature update, the raw feature of node i is concatenated with the mean-aggregated neighbor features, which results in a composite feature vector:

$$h_{i,concat}^{(0)}=CONCAT(h_i^{(0)},m_i^{(0)})$$

(5)

The concatenated features are linearly transformed using the layer L weight matrix W^(L) and the layer L bias term b^(L). Then, they are processed by the ReLU activation function to obtain the next layer of features:

$$h_i^{(1)}=\mathrm{Re}lu(W^{(0)}{h}_{i,concat}^{(0)}+b^{(0)})$$

(6)

All node features are arranged in node order to produce the first-layer node feature representation matrix:

$$H^{(1)}=\left[\begin{array}{c}{(h_1^{(1)})}^{\top }\\ ⋮\\ {(h_N^{(1)})}^{\top }\end{array}\right]$$

(7)

where N denotes the total number of nodes in the distribution network. H⁽¹⁾ has integrated the topological correlation information of the distribution network and node electrical features. Compared with the original electrical features, it can better characterize the local correlation changes in node states when a fault occurs, and it provides the input basis for the subsequent enhancement of fault-sensitive dimensions by SimAM.

2.3. SimAM Parameter-Free Attention Feature Enhancement

Once a fault occurs in the distribution network, the features of faulty nodes exhibit significant differences compared with those of normal nodes. The mean aggregation operation of GraphSAGE assigns equal weights to all features, thus failing to distinguish the importance of features in different dimensions for fault discrimination. SimAM defines an energy function to differentiate the importance of features in different dimensions for fault discrimination: the smaller the energy value, the greater the discrepancy between the node features and the global electrical background distribution, which in turn indicates a higher importance for fault discrimination.

On the node feature representation matrix H⁽¹⁾ output by the first layer of GraphSAGE, the set of feature representations of all nodes in the kth dimension is denoted as {h_1k,h_2k,…,h_Nk}. In the kth dimension, SimAM evaluates the anomaly of node features relative to the electrical background distribution. It takes the feature representation h_ik of node i as the target to be evaluated, while the feature values of the remaining nodes in this dimension are regarded as the global electrical background distribution under the current operating state of the distribution network. The mean and variance are used to represent the global background mean and dispersion degree of the node representations of all nodes in the kth dimension, respectively:

$${\mu }_k={\displaystyle \frac{1}{N}}\sum _{j=1}^N{h}_{jk}$$

(8)

$${\sigma }_k^2={\displaystyle \frac{1}{N}}\sum _{j=1}^N{(h_{jk}-{\mu }_k)}^2$$

(9)

Let λ denote the regularization coefficient. Then, the minimum energy of node i along the kth feature dimension can be expressed as

$$e_{ik}^{*}={\displaystyle \frac{4({\sigma }_k^2+\lambda )}{{(h_{ik}-{\mu }_k)}^2+2{\sigma }_k^2+2\lambda }}$$

(10)

The energy value is used to characterize the anomaly significance of node features relative to the global electrical background distribution under the current operating state. The term (h_ik − μ_k)² in the denominator of Equation (10) quantifies the degree of deviation of the feature representation h_ik of node i in the kth dimension from the background mean μ_k. When a single-line-to-ground fault occurs in the distribution network, the symmetric operating state of the three-phase system is disrupted. Key electrical quantities such as the fault-phase voltage and fault-phase current of the faulty node change, and such changes are reflected in the corresponding fault-sensitive dimensions, leading to an increase in the term (h_ik − μ_k)². The larger the deviation, the lower the energy value $e_i^{\ast }$, which indicates that the feature representation of node i in this dimension has higher importance for fault identification and location.

SimAM uses the reciprocal of the minimum energy as a fault-sensitive saliency measure in order to convert energy values into attention weights related to fault importance. Normalization processing is performed via the Sigmoid function, and the normalized result is subjected to element-wise weighting with the node feature representation matrix H⁽¹⁾, thereby yielding the enhanced feature representation matrix:

$$\tilde{H}=\mathrm{sigmoid}\left({\displaystyle \frac{1}{E}}\right)\odot H^{(1)}$$

(11)

where E represents the minimum energy value. SimAM enhances features using existing data and only taking into account the mean, variance and regularization coefficient of node feature representations, without introducing any trainable parameters.

Traditional attention mechanisms require training additional multi-layer perceptrons to generate weights. This introduces new parameters and can easily lead to overfitting in the presence of limited fault samples. Without introducing extra parameters, SimAM is less prone to overfitting and fully takes advantage of the following fact: the feature performance of a faulty node in key dimensions is changed, while the features of the vast majority of normal nodes remain relatively stable. Furthermore, before being fed into the next layer, all fault-sensitive dimensions are enhanced, which provides more discriminative input for the subsequent layer.

2.4. Multi-Head Attention Mechanism

SimAM realizes the enhancement of fault-sensitive feature dimensions and solves the weight assignment problem of different feature dimensions within the same node. To further distinguish the contribution of different neighbor nodes to the fault discrimination of the current node, this paper introduces a GAT layer after the SimAM module. The feature representations enhanced by SimAM have highlighted the fault-related feature dimensions, and GAT learns differentiated neighborhood weights according to the fault correlation between adjacent nodes and the current node. This enables neighbor nodes more related to the fault propagation path to obtain higher weights, while neighbor nodes with weaker correlation are assigned lower weights. The feature representations enhanced by SimAM are used as the input of the attention layer, which results in

$$h=\{h_1,h_2,\dots ,h_N\},h_i\in R^F$$

(12)

where N represents the total number of nodes in the distribution network, and F stands for the dimension of the node feature representations. For the central node i and its neighbor node j, the attention coefficient e_ij can be obtained through calculation:

$$e_{ij}=a(W{h}_i,W{h}_j)$$

(13)

The attention coefficient e_ij is used to measure the degree of contribution of neighbor node j to the fault discrimination of node i. To ensure the comparability of these coefficients across different neighbor nodes, GAT adopts the Softmax function to perform normalization processing on the attention coefficients between node i and its adjacent nodes. The attention weight α_ij is obtained.

$${\alpha }_{ij}={\displaystyle \frac{\exp (Leaky\mathrm{Re}lu(a^T[W{h}_i\parallel W{h}_j]))}{{\displaystyle \sum _{k\in N_i}}\exp (Leaky\mathrm{Re}lu(a^T[W{h}_i\parallel W{h}_k]))}}$$

(14)

where a represents the attention parameter vector and W stands for the feature transformation matrix. After the attention weight α_ij is obtained, the final output feature vector is updated as follows:

$$h_i^{\left(2\right)}=\sigma (\sum _{j\in N_i}{\alpha }_{ij}W{h}_j)$$

(15)

GAT employs a multi-head attention mechanism, which simultaneously characterizes the differences in fault correlation between neighbor nodes and the current node from multiple subspaces, thereby more fully capturing the neighborhood information related to fault discrimination. This mechanism utilizes K independent attention heads. The final output features can be obtained by concatenating the output results of these K groups. Below is the formula:

$$h_i^{\left(2\right)}=\underset{k=1}{\overset{K}{\parallel }}\sigma (\sum _{j\in N_i}{\alpha }_{ij}^k{W}^k{h}_j)$$

(16)

On the basis of the features enhanced by SimAM, the GAT layer performs differentiated weighted processing on neighborhood information, thereby highlighting the information of neighbor nodes more relevant to the fault propagation path and providing more fault-discriminative input for the subsequent deep aggregation of GraphSAGE.

2.5. GraphSAGE Deep Aggregation and Classification Output

After SimAM enhancement and GAT differentiated weighting, the output H⁽²⁾ of the second layer can already reflect the fault-related state differences between a node and its first-order neighborhood while it still mainly characterizes the first-order neighborhood information. In the final stage, GraphSAGE is reintroduced as the third layer to perform re-aggregation. Different from the differentiated neighbor weighting of GAT in the second layer, the function of GraphSAGE in the third layer is to integrate higher-order local topological information and comprehensively incorporate the influence of farther neighbor nodes along the fault propagation path on the fault state discrimination of the current node. The third layer of GraphSAGE adopts mean aggregation to sample and aggregate the neighborhood features output by the second layer. Third-layer features are obtained:

$$h_i^{(3)}=\mathrm{Re}lu(W^{(3)}{h}_{i,concat}^{(2)}+b^{(3)})$$

(17)

After re-aggregation by GraphSAGE, each node is ultimately represented as $h_i^{(3)}$. It integrates the node’s own electrical state, the weighted information of its first-order neighborhood, and higher-order local topological information. Therefore, $h_i^{(3)}$ provides a more complete feature representation for node-level fault discrimination.

To train model parameters, the problem of fault location is transformed into a problem of node classification, and the cross-entropy loss function is adopted:

$$L_{crossEntropy}=-\sum _i{y}_i\log {\widehat{y}}_i$$

(18)

where y_i represents the probability distribution predicted by the model, and ${\widehat{y}}_i$ stands for the true label. To prevent model overfitting, an L2 regularization term is introduced to constrain model complexity:

$$L_{L2}={\displaystyle \frac{\lambda }{2}}{‖W‖}_2^2$$

(19)

where λ denotes the regularization strength hyperparameter, and L2 represents the type of norm. Finally, the total loss function L of the model is a weighted sum of the L2 regularization term and the cross-entropy loss. It is defined as

$$L=L_{crossEntropy}+L_{L2}=-\sum _i{y}_i\log {\widehat{y}}_i+{\displaystyle \frac{\lambda }{2}}{W}_2^2$$

(20)

3. SimAM-GraphSAGE-GAT Integrated Distribution Network Fault Location Method

This paper constructs a SimAM-GraphSAGE-GAT architecture for distribution network fault location, as shown in Figure 2. This architecture is designed targeting the inherent characteristics of distribution network fault location, including local topological correlation, the imbalance of fault-sensitive feature dimensions, and the differences in contributions of different neighbor nodes. First, GraphSAGE is adopted to integrate the electrical information of each node and that of its first-order neighbor nodes and to generate initial node feature representations. Second, SimAM performs weight recalibration on feature dimensions to highlight fault-sensitive information, while GAT applies differentiated weighting to neighbor nodes according to the correlation between their fault states and that of the current node. Finally, a GraphSAGE layer is used to further aggregate high-order topological information, and node classification is performed to identify faulty nodes and normal nodes.

In this model, an undirected graph G is constructed based on the IEEE 33-node distribution network. Measurement data for active power, reactive power, three-phase voltage and three-phase current are concatenated to form the feature matrix X. The first layer employs GraphSAGE and performs neighbor sampling and local aggregation on each node. It preliminarily integrates the topological relationship and electrical features of the node and its first-order neighborhood and obtains the feature representation H⁽¹⁾. To solve the problem that the mean aggregation of GraphSAGE cannot distinguish the importance of different feature dimensions, H⁽¹⁾ is fed into SimAM to perform parameter-free recalibration on the features that contribute greatly to fault discrimination, thereby obtaining the enhanced feature representation. This step aims to enhance the fault-sensitive feature dimensions and reduce the influence of irrelevant background responses.

The second layer adopts the multi-head attention mechanism of GAT and takes the feature representations enhanced by SimAM as its input. This attention mechanism assigns different weights to adjacent nodes according to the correlation between their fault states and that of the current node, thereby obtaining the refined feature aggregation result H⁽²⁾. The multi-head attention mechanism can characterize the differences in fault correlation between neighbor nodes and the current node from multiple subspaces, thus more fully representing the fault-related neighborhood information.

The third layer adopts the GraphSAGE algorithm to aggregate high-order local topological information from the feature aggregation result H⁽²⁾ output by GAT. It outputs the fault category probability of each node through a fully connected layer and the Softmax function and classifies each node as a faulty node or a normal node, thus obtaining the final output result. The operation procedure of the SimAM-GraphSAGE-GAT model encoding is presented in Algorithm 1:

Algorithm 1. The Operation Procedure of the SimAM-GraphSAGE-GAT Model Encoding

Input: $A$, $X$, $W^{\left(0\right)}$, $W_c$, $W^k$, $W^{\left(3\right)}$, $b_c$, $b^{\left(0\right)}$, $b^{\left(3\right)}$

Output: $\widehat{Y}$

1. First GraphSAGE aggregation: $h_i^{\left(1\right)}=Relu(W^{\left(0\right)}{h}_{i,concat}^{\left(0\right)}+b^{\left(0\right)})$

2. SimAM Feature recalibration: $\tilde{H}=sigmoid\left({\displaystyle \frac{1}{E}}\right)\odot H^{\left(1\right)}$

3. GAT neighborhood weighting: $h_i^{\left(2\right)}=\underset{k=1}{\overset{K}{\parallel }}\sigma ({\displaystyle \sum _{j\in N_i}}{\alpha }_{ij}^k{W}^k{h}_j)$ 4. Second GraphSAGE aggregation: $h_i^{\left(3\right)}=Relu(W^{\left(3\right)}{h}_{i,concat}^{\left(2\right)}+b^{\left(3\right)})$ 5. Classification: $\widehat{Y}=Softmax(H^{\left(3\right)}{W}_c+b_c)$ 6. return: $\widehat{Y}$

4. Experimental Analysis

4.1. Model Construction and Data Generation

The IEEE 33-node distribution network model was built in MATLAB/Simulink (Version R2024b) based on an existing public template from MATLAB Central File Exchange and further modified to meet the fault-simulation requirements of this experiment [24]. The architecture is illustrated in Figure 3. The IEEE 33-node distribution network model consists of 32 lines. In total, 33 data measurement devices are installed in each node. The sampling frequency is 2 kHz. In the simulation, the fault time is in the range of 0.08–0.15 s, and the data are collected in the range of 0.05–0.15 s. Single-phase-to-ground faults are initiated on the lines, and fault resistance values are set to 0.1, 1, 7, 15 and 50 Ω, respectively. Fault samples are constructed by changing the location of faults and resistance values. Meanwhile, normal operating conditions are set as a control group. It is also necessary that the constructed dataset encompasses both fault and normal samples. For each operating condition, the measured values of active power, reactive power, three-phase voltage and three-phase current in the node are collected, and 200 samples can be obtained. Finally, 6400 samples can be obtained. To ensure that data under different operating conditions maintains the same proportions across different subsets, a stratified sampling method based on fault features is applied to the dataset. These samples are divided into a training set, a validation set, and a test set at a ratio of 8:1:1. Under this setting, more samples can be retained for model training while preserving enough samples in the validation set for hyperparameter tuning. The feature data is processed using the Min-Max normalization method. To prevent data leakage, the Min-Max normalization parameters are fitted solely on the training set, and the validation and test sets are normalized using the exact same parameters. Labels are assigned to each node in the sample. “0” means that the operating state is normal or “1” means that the operating state is a single-phase-to-ground fault.

4.2. Experimental Environment and Parameter Settings

Experiments are run on a Windows 11 ×64 operating system. The hardware configuration comprises an Intel Core i5-13500H central processing unit (CPU), 16 GB of random access memory (RAM) and an NVIDIA RTX 4060 graphics processing unit (GPU) (NVIDIA Corporation, Santa Clara, CA, USA). All the models are developed in PyTorch (Version 2.7.0). Loss rate, F1 score and accuracy are used as the evaluation metrics for the performance of the model.

Below are the formulas for precision and recall:

$$\mathrm{Pr}ecision={\displaystyle \frac{TP}{TP+FP}}$$

(21)

$$\mathrm{Re}call={\displaystyle \frac{TP}{TP+FN}}$$

(22)

The F1-score is computed as

$$\mathrm{F}1-score=2\times {\displaystyle \frac{\mathrm{Pr}ecision\times \mathrm{Re}call}{\mathrm{Pr}ecision+\mathrm{Re}call}}$$

(23)

Accuracy is the ratio of correctly predicted samples to the total number of samples in the dataset. The formula is given by

$$Accuracy={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$

(24)

where TP, FP, FN and TN denote true positive, false positive, false negative and true negative, respectively.

All models are uniformly configured with a three-layer structure in this experiment. The Adam optimizer is selected, and optimal parameters are chosen after multiple trials. Dropout is set as 0.3 for the purpose of suppressing overfitting. The hidden layer dimension is set to 64. The number of heads in the multi-head attention mechanism of the model is uniformly set to 4. The batch size and learning rate are set to 64 and 0.005, respectively. The number of iterations is set to 100. Weight decay is set to 0.0005. All experiments were repeated five times, and the results are reported as mean ± standard deviation.

4.3. Analysis of the Simulation Results

4.3.1. Comparison of Positioning Effects Across Different Models

The positioning results of the four models are demonstrated in

Table 2. Comparison of the positioning effects of four models.

Model	Loss	F1 Score (%)	Accuracy (%)	Training Time (s)
GraphSAGE-GAT	0.1135 ± 0.003	96.23 ± 0.24	96.76 ± 0.29	51.43 ± 0.1
GAT	0.2522 ± 0.012	85.81 ± 0.78	86.39 ± 0.82	65.77 ± 0.28
GraphSAGE	0.1825 ± 0.008	91.37 ± 0.51	91.86 ± 0.44	49.74 ± 0.06
SimAM-GraphSAGE-GAT	0.0969 ± 0.002	97.68 ± 0.13	98.24 ± 0.08	52.31 ± 0.09

, and positioning performance is displayed in Figure 4. The GraphSAGE-GAT model demonstrates higher accuracy and reduced fluctuation in comparison to GraphSAGE and GAT models. This reveals that incorporating the multi-head attention mechanism into the process of GraphSAGE sampling and aggregation enables the differentiated weighting of neighbor nodes. The contributions of different neighbor nodes to faults are represented in an accurate manner, which prevents information from critical fault nodes from being diluted during average aggregation and thereby enhances fault location capabilities.

After the integration of SimAM into the GraphSAGE-GAT model, the SimAM-GraphSAGE-GAT model further reduces loss rates and enhances F1 score and accuracy. The convergence curve is much smoother in the case of convergence. The reason is that SimAM uses an energy function for assessing the importance of information in feature dimensions and generating weights to recalibrate information in feature dimensions. The neighbor weights calculated by the GAT layer are made more stable by strengthening fault-sensitive features through feature recalibration without any extra parameters. Then, the model is enabled to focus more effectively on effective features and obtain better location performance.

Regarding training time, although the SimAM-GraphSAGE-GAT model requires an additional 0.88 s compared with the GraphSAGE-GAT model, its fault location performance is further improved. This is because SimAM does not introduce new parameters but uses existing data for feature recalibration. This improves performance without substantially increasing the training time of the model, which thus balances accuracy with efficiency.

4.3.2. Adaptive Analysis for High-Resistance Ground Faults

To evaluate the adaptability of the model under high-resistance faults, performance tests are conducted to verify the adaptability of the SimAM-GraphSAGE-GAT model when fault characteristics are attenuated. In the IEEE 33-node distribution network model, the fault point grounding impedances of 1000 and 1400 Ω are set. A fault location method based on ST-GCN was introduced as an external benchmark and tested under the same experimental settings to ensure a fair comparison. Samples are gathered for testing, and the experimental results are shown in

Table 3. Performance comparison of five models under the grounding impedances of 1000 and 1400 Ω.

Model	1000 Ω Grounding Impedance	1400 Ω Grounding Impedance
GraphSAGE-GAT	94.45% ± 0.19%	92.47% ± 0.26%
GAT	84.73% ± 0.64%	82.54% ± 0.59%
GraphSAGE	88.38% ± 0.42%	85.26% ± 0.48%
SimAM-GraphSAGE-GAT	96.49% ± 0.12%	95.71% ± 0.15%
ST-GCN	93.31% ± 0.26%	90.31% ± 0.19%

.

Under high-resistance faults, fault signals are weak. With the increase in ground fault impedance, all models exhibit a certain degree of performance degradation. Compared with the other compared models, the SimAM-GraphSAGE-GAT model achieves the best performance under the tested high-resistance grounding conditions. These results powerfully prove that SimAM can highlight feature dimensions that are more strongly correlated with faults, amplify major differences in dimensions, and make the subsequent weighted aggregation of neighborhoods more likely to form stable weights around effective information. As a result, the model can pay more attention to neighbor node information related to the fault, which hence improves its localization accuracy under high-resistivity faults.

4.3.3. Adaptive Analysis of Different Types of Noise and Data Loss Scenarios

Gaussian noise is chosen as the interference source and introduced into the test set to validate the adaptability of the SimAM-GraphSAGE-GAT model under various noise conditions. Gaussian noise levels are set to 20, 25 and 30 dB to simulate measurement interference and equipment errors.

Table 4. Performance comparison of four models under different noise disturbances.

Model	20 dB	25 dB	30 dB
GraphSAGE-GAT	90.47% ± 0.16%	92.53% ± 0.27%	94.17% ± 0.22%
GAT	78.83% ± 0.53%	81.34% ± 0.47%	83.12% ± 0.54%
GraphSAGE	83.51% ± 0.42%	86.74% ± 0.39%	89.31% ± 0.36%
SimAM-GraphSAGE-GAT	93.64% ± 0.17%	95.32% ± 0.13%	97.18% ± 0.11%

shows the experimental results. Under 20, 25 and 30 dB noise interference, the SimAM-GraphSAGE-GAT model achieves the highest accuracy among the compared models. This is attributed to the fact that limitations persist, though GraphSAGE smooths out Gaussian noise on local nodes to some degree by aggregating first-order neighbor features. Through the incorporation of SimAM, feature information across different dimensions is evaluated, and noise is mitigated at the feature dimension level, which thereby increases the signal-to-noise ratio of effective dimensions. This provides the GAT layer higher-quality feature data for the weighted processing of neighbor nodes, which facilitates the real-time perception of the connection status of the distribution grid. Consequently, the SimAM-GraphSAGE-GAT model effectively reduces the influence of noise on features and maintains stable fault location performance under the tested conditions.

To simulate the packet loss phenomenon caused by congestion or interruption of the communication network of the distribution network Supervisory Control and Data Acquisition (SCADA) system, in this paper, 3%, 6% and 9% of nodes are randomly selected for each sample in the test set, respectively, and all feature values of the selected nodes are set to zero to simulate the scenario of partial data loss.

As shown in

Table 5. Performance comparison of the four models under different data missing rates.

Model	3%	6%	9%
GraphSAGE-GAT	93.63% ± 0.24%	91.81% ± 0.2%	88.87% ± 0.28%
GAT	80.42% ± 0.61%	77.73% ± 0.59%	75.44% ± 0.56%
GraphSAGE	87.35% ± 0.37%	84.74% ± 0.41%	82.27% ± 0.39%
SimAM-GraphSAGE-GAT	96.28% ± 0.12%	94.7% ± 0.09%	93.53% ± 0.11%

, with the increase in the missing data rate, the fault location accuracy of all models decreases, indicating that the random absence of node features has an adverse impact on fault location. However, under different data missing rate scenarios, the SimAM-GraphSAGE-GAT model consistently maintains a high location accuracy. In contrast, GAT, which relies solely on the attention mechanism for neighborhood weighting, is more significantly affected by data missing. By aggregating the features of the node itself and its neighborhood, GraphSAGE possesses a certain capability for local information compensation. GraphSAGE-GAT combines the topological aggregation capability of GraphSAGE and the differentiated neighborhood weighting mechanism of GAT, improving the utilization efficiency of key neighbor node information. Based on this, SimAM-GraphSAGE-GAT incorporates fault-sensitive feature dimension recalibration, maintaining an accuracy of over 90% even at a 9% data missing rate, which indicates that this model has a stronger capability to resist data loss.

5. Conclusions

This paper addresses the issue of insufficient fault location performance of traditional fault location methods for distribution networks under complex operating conditions. It proposes a distribution network fault location method on the basis of SimAM-GraphSAGE-GAT. This method employs the SimAM-GraphSAGE-GAT model for the fault location of distribution networks. The following conclusions are drawn through simulation experiments and theoretical analysis:

(1)

This paper represents the distribution network topology as an adjacency matrix and integrates both the topological structure of the grid and electrical characteristic information. It converts the problem of fault location into a node classification task. SimAM is used to strengthen the information in the dimensions of important features to solve the problem of important feature information being attenuated after average aggregation. Finally, location accuracy is improved.

(2)

Under complex operating conditions involving varying noise interference, data loss and high-resistance grounding faults, the SimAM-GraphSAGE-GAT model consistently achieves a localization accuracy of over 92%. It outperforms the other models and shows good adaptability under the tested conditions.

(3)

GraphSAGE is employed as the first and last layer to efficiently sample and aggregate local and high-order neighborhood information. The parameter-free attention mechanism SimAM is introduced on top of GraphSAGE to enhance fault-sensitive features. In the meantime, the multi-head attention mechanism is applied as the intermediate layer to perform different weighting on neighbor nodes and thus shift more attention to fault-related nodes. This achieves an integrated design of the three components, which strikes a balance between accuracy and efficiency and improves fault location performance under the tested conditions.

This paper employs the SimAM-GraphSAGE-GAT model for fault localization. The model achieves high localization accuracy while ensuring computational efficiency on the IEEE 33-node distribution network. In the future, more fault-type data can be collected to strengthen the fault localization ability of the model in more complex and realistic scenarios by combining semi-supervised learning with transfer learning. In future work, more realistic noise models, including correlated PMU measurement errors, will be incorporated to further evaluate the robustness of the proposed method. The performance of the proposed method will also be assessed on larger and more comprehensive systems, such as the IEEE 123-node distribution network. In addition, the development of lightweight modeling methods may facilitate the deployment of the algorithm on terminal devices and support intelligent operation and maintenance in distribution networks.