A Power Grid Topological Error Identification Method Based on Knowledge Graphs and Graph Convolutional Networks

Abstract

Precise and comprehensive model development is essential for predicting power network balance and maintaining power system analysis and optimization. The development of big data technologies and measurement systems has introduced new challenges in power grid modeling, simulation, and fault prediction. In-depth analysis of grid data has become vital for maintaining steady and safe operations. Traditional knowledge graphs can structure data in graph form, but identifying topological errors remains a challenge. Meanwhile, Graph Convolutional Networks (GCNs) can be trained on graph data to detect connections between entities, facilitating the identification of potential topological errors. Therefore, this paper proposes a method for power grid topological error identification that combines knowledge graphs with GCNs. The proposed method first constructs a knowledge graph to organize grid data and introduces a new GCN model for deep training, significantly improving the accuracy and robustness of topological error identification compared to traditional GCNs. This method is tested on the IEEE 30-bus system, the IEEE 118-bus system, and a provincial power grid system. The results demonstrate the method’s effectiveness in identifying topological errors, even in scenarios involving branch disconnections and data loss.

1. Introduction

In the field of power system analysis and optimization, building precise and comprehensive models is essential for predicting and maintaining power network balance, which directly impacts the grid’s reliability and efficiency [1]. The rapid advancement of grid measurement systems and big data technologies poses significant challenges to grid security systems in terms of modeling, simulation, and fault prediction [2]. The increasing uncertainty in renewable energy use presents unprecedented challenges to the safe and cost-effective operation of power systems. Consequently, in-depth analysis and utilization of grid operational data have become essential to ensuring the safe and stable operation of power grids [3,4].

In this context, timely and accurate identification of grid topology and its changes using real measurement data has become crucial for data-driven grid operation and control. Network topology analysis is fundamental for enabling advanced power system management applications. Any errors in the topology can compromise data analysis in these applications and increase the likelihood of security analysis failure. Many identification methods, however, have low error tolerance, require extensive redundant information, and rely heavily on complete active power data.When data is missing or there are significant errors, the identified optimal data combinations will be directly affected [5]. The algorithm also needs a lot of redundant data to support it, and since power management units (PMUs) are not used very often right now, getting the extra data would cost a lot of money [6]. All of this adds to the algorithm’s complexity and implementation costs.

Traditional topological error identification methods include rule-based methods [7], residual methods [8], state estimation methods [9], and Artificial Neural Networks (ANN) [10,11], among others. The rule-based method is simple and fast to implement. It requires minimal computational and human resources when integrated into existing systems, and is reliable under various conditions. However, it is less effective for identifying plant-station topological errors [7]. Residual methods are straightforward after state estimation, but when both analog bad data and topological errors are present, the residuals will exceed the threshold, making it difficult to distinguish between analog and topological errors [8]. Traditional state estimation methods establish a relationship between measurement residuals and topological branches using the measurement branch association matrix, applying covariance principles to identify topological errors [9]. In addition to traditional state estimation, Generalized State Estimation (GSE) is widely used for topological error identification [12]. This method integrates network topology and system parameters, and the algorithm converges efficiently. However, GSE faces challenges with the large number of analyzed nodes, high dimensionality of the iteration matrix, and slow computational speeds. Recently, Wu et al. treated the measurement data as a non-smooth Gaussian process, detecting errors by testing whether the mean vector of the process is zero, independent of the convergence of the standard weighted least squares (WLS) state estimation algorithm [13]. This method addresses the issue of non-convergence in traditional state estimation algorithms and is highly effective in detecting topological errors. These methods do not require extensive historical data for modeling, and have transparent, easily understandable, and verifiable algorithmic steps, providing new solutions and advances in grid topological error identification. However, traditional methods lack flexibility in handling complex and dynamic grid structures and are inefficient in processing large-scale datasets [14].

In recent years, deep learning methods have been extensively applied to power grid topology identification, with commonly used models including convolutional neural networks (CNN), deep neural networks (DNN), and graph convolutional networks (GCN). Wang et al. employed artificial neural networks and power grid balance equations for the identification of grid topological errors [15]. By treating the measured residual state estimation as the optimization objective and the identification outcomes of the state estimation as the input, the neural networks were repeatedly trained to identify switching state errors and electrical connection errors more quickly. Tian et al. employed CNN for smart grid topology identification, which aims to address the issues of frequently changing power grid topology and challenges in learning deep features of data [16]. Additionally, Zhong et al. formed a CNN-LSTM hybrid neural network by combining the spatial feature extraction capability of CNN and the temporal feature extraction capability of LSTM [17]. This allowed them to finally classify the samples to obtain the topology relationship, thus improving the accuracy of topology identification. Through the introduction of deep learning and neural network approaches, these studies significantly advance the development of smart grid topology identification technology [18].

However, deep learning methods are less frequently applied in topological error identification. Traditional artificial neural networks, although generally effective in topological error identification, require extensive training samples, limiting their practical application in real systems. As deep learning models become widely adopted, traditional artificial neural networks struggle to meet the requirements for topological error identification in modern large-scale power grids. While CNN, DNN and similar networks typically handle data with Euclidean characteristics and are restricted to spatial and local feature extraction, this limitation makes it challenging to process non-Euclidean data and capture relationships between nodes, leading to insufficient exploration of the network structure and node connections. In contrast, graph convolutional networks (GCNs) are well-suited for processing non-Euclidean data, enabling the identification of dynamic topological errors in power grids through graph convolution operations [19].

GCNs offer a broad range of applications in grid engineering and power systems, including grid topology identification, state estimation, and fault diagnostics [20,21]. For instance, Zhang et al. introduced GCNs into fault diagnosis to address the limitations of traditional machine learning methods [22]. They employed GCNs for node information aggregation and wavelet transforms to analyze time-domain fault data, achieving higher accuracy and robustness than CNN and DCN. In fault recognition, Li et al. proposed a novel approach to address the challenges of low accuracy and poor real-time performance in existing methods [23]. They modeled grid topology as line graphs and trained GCNs using node features. Therefore, based on the insights from these studies, applying GCNs to topological error identification is feasible.

Knowledge graphs are increasingly applied across various industries due to their ability to clearly and flexibly represent relationships between concepts, entities, and their attributes. For instance, Feng et al. utilized the knowledge graph’s rapid analysis capabilities and robust semantic processing to develop a fault operation and maintenance knowledge graph (FOM-KG) within the power information collection system [24]. This significantly improved the efficiency of operation and maintenance processes. Many researchers combine knowledge graphs with GCNs, as the graphical nature of knowledge graphs aligns with GCNs’ data processing requirements. To address the limitations of using a single knowledge graph approach, Wang et al. applied both knowledge graphs and GCNs for topology identification, using GCNs to analyze relationships between elements identified through the knowledge graph [25]. Inspired by this study, we constructed a knowledge graph for grid data, which was subsequently used with GCNs to identify topological errors.

The topology of a power system represents the relationships between nodes in graphical form, making the knowledge graph an ideal tool for representing this topological information. However, topological errors can arise in the network due to events like line disconnections or data loss. GCNs, designed to analyze graph data and explore relationships between entities in a knowledge graph, can effectively learn the connections between nodes and edges to detect potential topological errors. Thus, the characteristics of knowledge graphs and GCNs make them well-suited for identifying topological errors.

This paper further investigates the combined application of the Knowledge Graph and Graph Convolutional Network (GCN), building on the work of previous researchers. The main contributions of this paper are as follows:

• We propose a knowledge graph construction method for power grids that transforms nodes (e.g., generators, transformers, lines) and their relationships into a knowledge graph, offering rich contextual information for topology analysis.
• To address the limitations in information aggregation with traditional GCNs, we propose a new GCN model incorporating mechanisms like multilayer graph convolution, neighborhood learning, and splicing operations. This model enhances the ability to capture both node and neighborhood features while preserving node-specific information, thus improving model flexibility.
• By combining the knowledge graph with the improved GCN, we input the processed knowledge graph data into the GCN for training. The GCN can identify potential topological errors, such as line disconnections or data loss, by learning node and edge features. Test results demonstrate that this method enhances identification accuracy and robustness compared to traditional GCN–knowledge graph combinations, offering an effective solution for intelligent topological error identification in larger-scale power grid systems.

2. Methodological Principles

2.1. Knowledge Graph

A knowledge graph uses graph models to describe and represent various concepts and their relationships. It depicts complex relationships and attributes, while providing an efficient way to store, retrieve, and manage information by organizing these elements in a structured form. Additionally, it supports complex queries and data analysis operations [26]. The concept of knowledge graph originated from the Semantic Web at the earliest [27], and its goal is to transform the traditional text-based World Wide Web into the entity-based Semantic Web, allowing the knowledge graph to express real-world knowledge more accurately.

Academic definitions of knowledge graphs, although somewhat ambiguous, generally describe their semantic representation through formalization. For instance, Ehrlinger et al. state that “A knowledge graph acquires and integrates information into an ontology and applies a reasoner to derive new knowledge” [28]. This perspective emphasizes the dynamic nature of knowledge graphs and the capacity to derive knowledge. A different perspective suggests that “A knowledge graph is a multi-relational network composed of entities and relations, which are conceptualized as nodes and different types of edges, respectively” [29]. It underscores the structural nature of knowledge mapping, where complex knowledge systems are built by combining entities and relationships.

A knowledge graph is essentially a complex semantic network, with its basic data structure summarized as a triad: ‘entity-relationship-entity’ or ‘entity-attribute-attribute value’ [30]. In this structure, entities represent objects, events or attributes in the real world, and serve as the core elements in the knowledge graph. Relationships connect two entities and express some kind of semantic connection between them, while attributes describe entities, acting as a special relationship between entities and attribute values, also stored in the triad. Typically, the knowledge graph formal structure is defined as:

$$G=E\times R\times E=\left\{(h,r,t)|h,t\in E,r\in R\right\}$$

(1)

where E represents the entity ensemble, R represents the relation ensemble, $(h,r,t)$ is a ternary, h, r, t denote the head entity, relation, and tail entity, respectively.

The schema layer and data layer, whose hierarchical topologies are shown in Figure 1, form the two main components of a knowledge graph [31]. The schema layer forms the logical foundation of a knowledge graph, which defines the types and structures of different data items. The schema layer delineates the entities, relationships, attributes, and other elements while defining the constraints and hierarchical structures. In the schema layer, entity concepts are typically represented by nodes, with edges depicting the relationships between them. In the data layer, specific information is stored in triple form, where nodes represent unique entities and edges indicate relationships between entities or between qualities and entities. The integration of the schema and data layers enables the knowledge graph to store and manage vast amounts of data.

2.2. Graph Convolutional Network

2.2.1. Basic Concepts of Graph

A graph is a data structure used to represent complex relationships between different objects. It consists of two components: edges, which connect nodes and indicate relationships between them, and nodes, which serve as the fundamental elements representing objects or entities [32]. A simple graph can be represented by the following equation:

$$G=\{V,E\}$$

(2)

where V represents the set of nodes and E represents the set of edges. A graph is called directed if all of its edges have a direction; otherwise, it is considered undirected. An edge consists of two nodes. A simple graph G is illustrated in Figure 2.

For instance, in Figure 2, edges in the set E are represented by the notation $e_{ij}=(v_i,v_j)$, which indicates the edges connecting nodes $v_i$, $v_j$. Nodes in the set V are denoted as $v_i$, $v_j$.

In the graph $G=\{V,E\}$, an adjacency matrix $A\in R^{n\times n}$, where n is the number of nodes [33], is typically used to indicate connectivity relationship between the nodes. As shown in the equation below, each element in A is denoted as $a_{ij}$:

$$a_{ij}=\left\{\begin{array}{cc}1,\hfill & v_i\to v_j\hfill \\ 0,\hfill & otherwise\hfill \end{array}\right.$$

(3)

where: $a_{ij}$ represents the element in the i-th row and j-th column of the adjacency matrix A. If an edge exists between node i and j, then $a_{ij}=1$, and otherwise, then $a_{ij}=0$.

Generally, adjacency matrices are represented as two-dimensional arrays, while node set relations are depicted as one-dimensional arrays. If G is an undirected graph, then A is a symmetric matrix, and otherwise, A is not necessarily symmetric. Taking $n=5$ as an example, the adjacency matrix of the graph is shown as

$$A=\left[\begin{array}{ccccc}0& 1& 1& 0& 0\\ 1& 0& 1& 1& 0\\ 1& 1& 0& 1& 0\\ 0& 1& 1& 0& 1\\ 0& 0& 0& 1& 0\end{array}\right]$$

(4)

Nodes can also be represented using the feature matrix X and the metric matrix D. The feature matrix X is an $N\times f$ matrix, where N represents the number of nodes and f denotes the dimension of the feature vector. Each node’s features can include any relevant information, such as attributes or statistics. The metric matrix D is shown as

$$D=\left[\begin{array}{ccccc}2& 0& 0& 0& 0\\ 0& 3& 0& 0& 0\\ 0& 0& 3& 0& 0\\ 0& 0& 0& 3& 0\\ 0& 0& 0& 0& 1\end{array}\right]$$

(5)

the dimension of the metric matrix D is $N\times N$. The diagonal element $d_{ii}$ denotes the degree of node i, representing the number of edges connected to node i, calculated by the following equation.

$$D_{ii}=\sum _{j=1}^n{a}_{ij}$$

(6)

2.2.2. Basic Principles of GCN

The Graph Convolutional Network (GCN) is a type of graph neural network commonly used for tasks such as node classification and link prediction. The core idea of GCN is to learn the node representations and capture information from the local neighborhood through convolutional operations on graph structures, enabling the handling of non-Euclidean data that traditional deep learning methods struggle to process directly. For data with graph structure, GCN aggregates and transforms node and neighboring node information by stacking multiple convolutional layers [34], the specific steps include:

• Feature Aggregation (Aggregation): each node collects the features from its neighbors;
• Feature Update (Update): each node updates both its own and its neighbors’ features based on the aggregated information, forming a new node representation.

Both spatial-domain and spectral-domain based GCNs are examples of current GCNs [35]. Spectral-domain GCNs, based on spectral graph theory, use graph signal processing filters to construct graph convolutions, which help eliminate noise in graph signals and offer a convolutional neural network formulation.

This paper primarily employs spatial-domain GCN, which operates directly on the spatial structure. Specifically, each node computes the weighted average of its neighbors’ features, combines them with its own, and applies an activation function to generate a new feature representation. Figure 3 illustrates the convolution process for a single node in the spatial convolution of a GCN.

As shown in Figure 3, nodes 2, 3, 5 and 7 (including node 4 itself) can be used to convolve node 4, which can be obtained as $n_4$. Thus, the convolution of node 4 $Con{v}_4$ can be expressed as:

$$Con{v}_4=W_{2,4}{v}_2+W_{3,4}{v}_3+W_{5,4}{v}_5+W_{7,4}{v}_7+W_{4,4}{v}_4$$

(7)

where v denotes the features of each node and W denotes the weights of the convolution.

The graph $G=(V,E)$ consists of V, the set of nodes, and E, the set of edges. The input to the initial layer is $H^{\left(0\right)}=X,X\in R^{N\times d}$. The adjacency matrix is $A\in R^{N\times N}$, where N represents the number of nodes. The graph convolution operation at the l-th layer can be expressed as:

$$H^{(l+1)}=\sigma \left(A{H}^{\left(l\right)}{W}^{\left(l\right)}\right)$$

(8)

where $H^{\left(l\right)}$ represents the input feature matrix of the l-th layer, $H^{(l+1)}$ represents the output feature matrix of the $l+1$-th layer, $W^{\left(l\right)}$ is the weight matrix from the l-th layer, and $\sigma$ is the activation function. GCN aggregates the neighboring features of a node by multiplying the adjacency matrix A and the feature matrix H [36].

Since the previous formulas ignore the influence of a node on its own features, the Laplacian matrix $L=D-A$ is introduced, where $D\in R^{N\times N}$ is the degree matrix. The formula for the l-th convolutional layer is now:

$$H^{(l+1)}=\sigma (D^{-{\displaystyle \frac{1}{2}}}L{D}^{-{\displaystyle \frac{1}{2}}}{H}^{\left(l\right)}{W}^{\left(l\right)})$$

(9)

where $D^{-{\displaystyle \frac{1}{2}}}$ is the matrix obtained by taking the inverse square root of the degree matrix D [37]. This formula incorporates the Laplace matrix, the degree matrix, and the normalizing operation to solve the problem of closing the loop on the characteristics of its own nodes. The structure of the GCN is schematically shown in Figure 4.

2.3. Topological Error Identification

The topology of a power grid defines the connections between its components, and topological error identification determines the presence of errors by analyzing the states and connections of these components. Traditional topological error identification techniques are insensitive to changes in input data quality. In contrast, GCNs can model the grid topology as a graph. By training on topology data, GCNs can learn to recognize normal and abnormal states in the network structure [25]. This work presents a novel grid topological error identification method that combines knowledge graph with GCN, leveraging the fault-tolerant characteristics of GCNs. Traditional knowledge graphs can efficiently produce graph-structured data and extract potential relationships between entities, but their accuracy is limited when the data contains errors or conflicting information. On the other hand, GCNs possess robust fault-tolerance mechanisms, enabling them to handle errors and conflicts in graph data.

In this paper, we treat entities in the knowledge graph as “nodes” and the relationships among them as “edges”. Then we utilize historical topological data and state information to train the GCN, enabling it to learn normal and abnormal patterns in the topology, and detect potential errors. This method effectively addresses the limitations of using knowledge graphs alone by integrating GCN, offering a novel and efficient solution for precise grid topological error identification. The particular flowchart is shown in Figure 5.

3. Methodology of the Proposed Method

3.1. Knowledge Graph Construction

The knowledge graph construction process is illustrated in Figure 6.

Semi-structured data is used to construct the knowledge graph, with information extracted using specific wrappers for formats like HTML and XML.

The grid steady-state model data from the State Grid Cloud is primarily presented in the form of Excel sheets containing semi-structured data. Python is used to extract necessary data from the Excel sheets, including values such as the start point, end point, first-end active, and first-end reactive power, as shown in Figure 7. The data is then organized into a ternary entity-attribute structure, forming an array with elements such as: branch, power status at the endpoint, corresponding attribute value; branch, start point, node serial number, etc. Finally, the ternary form of the knowledge graph is saved in a JSON file for later error analysis and optimization.

After information extraction, the data enters the knowledge fusion phase, which includes entity disambiguation and co-reference resolution. These processes aim to eliminate the ambiguity of entity references and improve the accuracy of entity associations and referential relationships, ensuring more precise semantic interpretations for knowledge mapping tasks.

Following knowledge integration, a series of knowledge processing steps, including ontology construction, are performed. Ontology construction entails designing and building the ontological framework of the knowledge graph. In this context, an ontology refers to a formal expression for describing the relationships between concepts and entities in a particular domain, which defines terms, classifications, attributes, and relationships within the domain and specifies the hierarchies and constraints between them.

Data Intelligence Diagnostics play a crucial role in the construction and maintenance of the Knowledge Graph. This task is a comprehensive and automated analysis of the data in the knowledge graph, with the aim of revealing data quality issues, potential errors, and inconsistencies, and providing appropriate adjustments.

This study introduces the visualization function, incorporating an intelligent search and recommendation engine, aimed at improving the operability and user experience of the knowledge graph. The knowledge graph of power grid data is constructed and visualized through the following steps, utilizing the theory of infinite automata and the data visualization tool Neo4j: First, the power grid data are extracted from the Excel sheet. Next, the infinite automata are employed to extract keywords from the data, which are then organized into triples and imported into the Neo4j database for visualization. The visualized knowledge graph provides detailed node information, including node number, reactive load estimate (WLS voltage magnitude), node name, reference voltage, and node type.

3.2. GCN Structure

The GCN structure of this paper is shown in Figure 8.

After processing the knowledge graph data, GCN processes the graph structure data by jointly learning node features and the adjacency matrix. This method enhances information extraction, and increases the model’s efficiency in learning deeper graph structure features.

The GCN accepts two main inputs: node feature matrix and adjacency matrix. The Node Feature Matrix captures the features of each node, while the Adjacency Matrix defines the connectivity relationship among the nodes [38,39]. In GCN, the update of each node depends on its own features and those of neighboring nodes weighted by the adjacency matrix.The result of neighborhood aggregation can thus be represented as:

$$output=AM\otimes FM$$

(10)

where $AM$ represents the adjacency matrix, $FM$ denotes the feature matrix, and $output$ is the result of neighborhood aggregation.

To mitigate information loss during the convolution process, this paper introduces concatenation and employs a three-layer graph convolution. The graph convolution process for each layer is as follows:

$$out\left[i\right]=ReLU\left(BN\left(LN(x_i\oplus outpu{t}_i)\right)\right)$$

(11)

where $out\left[i\right]$, $x_i$ and $outpu{t}_i$ represent the output, input, and neighborhood aggregation result of the i-th layer, respectively.

During graph convolution, the graph convolution layer linearly transforms node features into a new feature space, where the adjacency matrix aggregates the transformed features from neighboring nodes to achieve feature fusion. Compared to the traditional GCN, the model in this paper retains feature splicing before and after aggregation, which better preserves the original node information and addressing the issue of information loss during convolution. Moreover, the additional linear and BN layers improve the model’s learning ability and enhance the flexibility and adaptability in the topological error identification.

3.3. Loss Function

The loss function used in the network training process is the Cross-Entropy Loss Function [40,41], which is formulated as follows:

$$Loss=\sum _{m=1}^N[y_{i}log{\widehat{y}}_i+(1-y_i)(1-log{\widehat{y}}_i)]$$

(12)

where $Loss$ is the total loss of all nodes, $y_i$ is the true value of the i-th node, and ${\widehat{y}}_i$ is the observed value of the i-th node. Finally, in the output section, the feature matrix output by the network is used to represent the classification result corresponding to each node, i.e., the final topological error identification result.

4. Experiment

4.1. Experimental Datasets

This paper compares the experimental results of binary and ternary classifications. Specifically, binary classification involves two cases: normal lines and disconnected lines. The comparison between actual and predicted disconnected branches determines the accuracy of the binary classification. Ternary classification involves three cases: normal lines, disconnected lines and lines with data loss. The data loss, including parameters such as resistance, inductance, capacitance, and branch ratio, can occur randomly. This paper also compares the actual and predicted disconnected lines, as well as the actual and predicted data loss branch, to determine the accuracy of the ternary classification.

4.1.1. Set-Up of IEEE 30-Bus System

The IEEE 30-bus system is a widely used small-sized power system test model containing 30 nodes and 41 branches, including buses, transformers, lines, generators, and other components. The system is frequently utilized in power system studies such as trend analysis, fault simulation, and economic dispatch. The topology of this system is shown in Figure 9.

In the binary classification experiment of IEEE 30-bus system, two lines (2-30, serial number 15, and 11-28, serial number 9) were selected to compare actual and predicted disconnected lines, yielding binary classification accuracies. In the ternary classification experiment, two lines are randomly disconnected for each sample (2-30, serial number 15, and 11-28, serial number 9) along with 8 randomly selected lines that miss certain attribute values.

4.1.2. Set-Up of IEEE 118-Bus System

The IEEE 118-bus system, a widely used standard test model in power system research, contains 118 nodes, 54 generators, and 186 branch circuits that simulate a complex power network. The node data includes information on voltage magnitude, phase angle, load and generation, while the system circuit data includes resistance, reactance and transmission capacity. The IEEE 118-bus system is extensively used in research areas like power system trend analysis, optimal dispatch and fault detection. The topology of this system is shown in Figure 10.

In the binary classification experiment of IEEE 118-bus system, two lines (8-30, serial number 37, and 48-49, serial number 69) were selected to compare actual and predicted disconnected lines, yielding binary classification accuracies. In the ternary classification experiment, two lines are randomly disconnected for each sample (8-30, serial number 37, and 48-49, serial number 69) along with 8 randomly selected lines that miss certain attribute values.

4.1.3. Set-Up of Grid Data of a Province

The grid data of a province contains network topology parameters and measurement data, with a total of 1389 nodes, 2419 branches, and 9974 measurements. The data setup is comparable. In the binary classification experiment, this paper selects two lines (798-34, serial number 596 and 184-211, serial number 2258). In the ternary classification experiment, two lines are randomly disconnected (798-34, serial number 596 and 184-211, serial number 2258), along with 180 randomly lines to miss certain attribute values.The structure of the grid data is shown in Figure 11.

4.2. Experimental Environment

The experimental tests was conducted in the Pycharm environment, with the computer hardware specifications including a 13.3 GHz 13th Gen Intel(R) Core(TM) i9-13900HX CPU and an NVIDIA Geforce RTX 4060 laptop GPU. Python 3.10 and the Pytorch 2.0.0 toolkit were used with the deep learning framework to implement the topological error identification model. The specific parameters of this paper are listed in

Table 1. Parameter Setup.

Parameter Name	Value
Number of graph convolution layers	3
Activation function	ReLU
Optimizer	AdamW
Loss	Cross-Entropy Loss
Epoch	5000
Learning rate	0.01

.

4.3. Experimental Results and Analysis

The binary and ternary classification results for grid topological error identification, averaged over ten experiments, are presented in

Table 2. Accuracy of Binary and Ternary Classification on IEEE 30.

Accuracy of Binary Classification	Accuracy of Ternary Classification
95.61%	83.90%

for the IEEE 30-bus system.

The binary and ternary classification results for grid topological error identification, averaged over ten experiments, are presented in

Table 3. Accuracy of Binary and Ternary Classification on IEEE 118.

Accuracy of Binary Classification	Accuracy of Ternary Classification
99.98%	96.53%

for the IEEE 118-bus system.

The method is also tested on the provincial power grid data and the experimental results are presented in

Table 4. Accuracy of Binary and Ternary Classification on Provincial Power Grid Data.

Accuracy of Binary Classification	Accuracy of Ternary Classification
99.92%	99.21%

.

The results demonstrate that the method in this paper accurately identifies disconnected lines and data loss conditions on both the IEEE 118-bus and provincial grid data, with an accuracy exceeding 95%. For the IEEE 30-bus system, the method achieves high accuracy in the binary classification of disconnected lines but is less effective in the ternary classification of data loss compared to the other systems. However, the accuracy remains within an acceptable range, considering the limited topology information in the IEEE 30-bus system. Complex models like GCNs may struggle to fully capture topological features in smaller systems. In contrast, the more complex structure and larger neighborhoods in the IEEE 118-bus and provincial grids enable GCNs to learn from extended neighborhood information, enhancing the model’s ability to identify topological errors.

In this paper, the GCN model was compared to the traditional GCN models [42] for experiments, and the accuracy was averaged over ten experiments. The specific results are presented in

Table 5. Results of GCN Proposed Compared To traditional GCN on IEEE 30.

	Accuracy of Binary Classification	Accuracy of Ternary Classification
Proposed	95.61%	83.90%
Conventional GCN	95.12%	81.95%

,

Table 6. Results of GCN Proposed Compared To traditional GCN on IEEE 118.

	Accuracy of Binary Classification	Accuracy of Ternary Classification
Proposed	99.98%	96.53%
Conventional GCN	99.98%	94.93%

and

Table 7. Results of GCN Proposed Compared To traditional GCN on Provincial Power Grid Data.

	Accuracy of Binary Classification	Accuracy of Ternary Classification
Proposed	99.92%	99.21%
Conventional GCN	99.91%	97.05%

.

Compared to the traditional GCN, the GCN proposed in this paper shows varying degrees of improvement on both the IEEE 118 system and provincial grid datasets, particularly in the ternary classification task of the provincial grid data. The two models show minimal differences in binary classification results on the IEEE 30 system, while the proposed model demonstrates improvements in ternary classification with data loss. These improvements are attributed to the enhanced feature processing flow in the proposed GCN, which optimizes feature extraction and processing capabilities. The proposed GCN incorporates feature information during aggregation. Additionally, the added linear and batch normalization (BN) layers enhance the learning ability of the network, increasing its flexibility and adaptability in topological error identification. The splicing and transformation operations in the improved GCN resemble residual connections in residual networks, effectively mitigating gradient explosion and vanishing issues, while extending feature representation dimensionality. This is particularly significant for handling complex graph-structured data like power grid topological error identification.

5. Conclusions

In summary, this paper proposes a method for identifying topological errors by combining a knowledge graph with a GCN. First, the construction of the knowledge graph transforms grid data and their relationships into a graph format, effectively managing complex data and providing valuable insights for topology analysis. Additionally, this paper proposes a new GCN model that improves the ability to capture both node and neighborhood features while preserving node-specific information. This is achieved through mechanisms such as multi-layer graph convolution, neighborhood learning, and a splicing operation. The processed knowledge graph data is then fed into the GCN for training, enabling the identification of potential topological errors, such as line disconnections or data loss. Testing on an IEEE 30-node system, an IEEE 118-node system, and data from a provincial power grid demonstrates that this method improves both identification accuracy and robustness compared to traditional GCN-knowledge graph combinations, proving its effectiveness. This approach provides strong technical support for applications such as monitoring, optimal scheduling, and fault prediction in grid systems, and offers a scalable solution for intelligent grid topological error identification.