Power transformer health state intelligent assessment method based on knowledge graph and graph neural network

Abstract

The invention discloses a power transformer health state intelligent assessment method based on a knowledge graph and a graph neural network, and the method comprises the steps: carrying out the preprocessing of multi-source state data in the operation of a transformer, converting the data into a health score value, marking a health grade label, and obtaining a training data set; constructing a health state knowledge graph by using the training data set; based on the training data set and the health state knowledge graph, calculating an edge weight by adopting a graph attention network; based on the edge weight, utilizing a node feature aggregation mechanism guided by the edge weight to iteratively update the node embedding representation, and obtaining a node representation fusing the structure information and the feature information; and introducing an incremental learning mechanism on the health state knowledge graph, and updating the health state knowledge graph based on the multi-source state data at the current moment. According to the method, multi-source state knowledge can be fused, the graph structure expression ability is achieved, dynamic updating is supported, the structure expression ability is achieved, and edge weight self-adaptive updating and sustainable evolution are supported.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of based on knowledge graph and graph neural network's electric power transformer health state intelligent evaluation method, belong to electric power equipment state evaluation technical field. BACKGROUND

[0002] With the continuous expansion of the scale of power system and the continuous improvement of the proportion of new energy access, the operation environment of electric power equipment is increasingly complex, and higher requirements are put forward for the accuracy and timeliness of its health state evaluation. As a core component of power grid, the running state of electric power transformer is directly related to the safety and stability of the whole system. During the operation of transformer, it is affected by the coupling effect of multiple physical fields such as electricity, heat, machinery and chemistry, and long-term accumulation can easily cause potential faults such as winding deformation, core loosening and insulation aging, which will affect the normal operation of the equipment, and even cause system failure or power interruption, causing serious economic loss and safety risk.

[0003] At present, the traditional state evaluation method mainly depends on expert rules, threshold judgment or single index analysis, which is difficult to fully reflect the correlation between multiple sources of transformer information, and is slow in response when facing dynamic data changes, with insufficient evaluation accuracy, which is difficult to meet the demand of modern power system for intelligent operation and maintenance. Especially in the face of large amount of operation data, complex coupling between characteristics and nonlinear state evolution, the existing method still has obvious shortcomings in the stability, sensitivity and explainability of evaluation results.

[0004] Traditional data-driven models often ignore the potential causal structure between features, cannot effectively express the linkage mechanism between multiple physical fields of transformer, and lack the ability of structured modeling and reasoning of state knowledge, making it difficult to realize accurate identification and timely warning of complex operating state. SUMMARY

[0005] The technical problem to be solved by the present application is to overcome the defects of the prior art and provide a kind of based on knowledge graph and graph neural network's electric power transformer health state intelligent evaluation method, the main purpose is to solve the problems of insufficient data structure expression, weak correlation modeling of key indicators and lag of evaluation result updating in transformer operating state evaluation in the prior art, which can integrate multi-source state knowledge, have graph structure expression ability and support dynamic updating, have structure expression ability, support edge weight adaptive updating and can evolve continuously.

[0006] Preferably, the present application provides a kind of based on knowledge graph and graph neural network's electric power transformer health state intelligent evaluation method, comprising: step 1: pre-processing the multiple source state data in the operation of transformer collected in advance, obtaining the per unit value, and constructing a standardized data set based on the per unit value;

[0007] Step 2: Convert the per-unit values ​​in the standardized dataset into health score values, label the health score values ​​with health level labels, and obtain the training dataset;

[0008] Step 3: Using the training dataset obtained in Step 2, construct a health status knowledge graph, with status indicators as nodes and physical or statistical associations as edges; the graph structure of the health status knowledge graph is G = (V, E), where V is the set of nodes and E is the set of edges;

[0009] Step 4: Based on the training dataset obtained in Step 2 and the health status knowledge graph constructed in Step 3, the edge weights are calculated using a graph attention network;

[0010] Step 5: Based on edge weights, use the edge weight-guided node feature aggregation mechanism to iteratively update the node embedding representation to obtain a node representation that integrates structural information and feature information;

[0011] Step 6: Introduce an incremental learning mechanism to the health status knowledge graph, and update the health status knowledge graph based on the multi-source status data collected at the current moment.

[0012] Prior to, step 1 includes:

[0013] A multi-dimensional transformer operation characteristic index library is constructed, which includes electrical parameters, thermal parameters, and mechanical parameters. Electrical parameters include load current and voltage fluctuations; thermal parameters include top oil temperature and ambient temperature; and mechanical parameters include core and tank vibration acceleration. Reference and optimal ranges for electrical, thermal, and mechanical parameters are determined. If an electrical, thermal, or mechanical parameter is outside the reference range, it is marked as abnormal. If the number of times an abnormal parameter is marked exceeds a preset threshold, the abnormal electrical, thermal, or mechanical parameter is removed and replaced with a robust interpolation value. The electrical, thermal, and mechanical parameters are normalized to per-unit values ​​to obtain a standardized dataset.

[0014] Prior to this, step 2 includes:

[0015] Calculate the health score S of the i-th per-unit value. i :

[0016]

[0017] Among them, S i X represents the health score value of the i-th per-unit value. i x represents the measured value of the i-th per-unit value in the per-unit values. i,opt Let x represent the optimal operating value of the i-th per-unit value. i,min k represents the minimum acceptable value for a per-unit value. i This is the score decay coefficient;

[0018] Based on attributes, health scores are categorized directionally, and weighting factors are assigned.

[0019] Based on pre-defined labels and historical data of abnormal operation and maintenance, health level labels are labeled for health score values ​​to obtain a training dataset.

[0020] Prior to this, step 3 includes the following steps:

[0021] Step 31: Using health score as the node candidate set, analyze the features in the training dataset, calculate the lagged correlation, mutual information index and Granger causality coefficient of the features, and construct the correlation matrix between features;

[0022] Step 32: For candidate nodes v in the node set V i With candidate node v j Perform a two-way test; if the correlation index of the candidate node is greater than the set threshold θ... r Furthermore, if the causality test is passed, then at candidate node v i With candidate node v j Constructing a directional edge e between them ij Obtain the optimized edge set and initialize the edge weights. As for the edge e ij Standardized relevant values;

[0023] Step 33: Extract the corresponding running feature vectors for all nodes in the graph structure G = (V, E). R d For a real space of dimension d, for the eigenvectors Perform normalization and dimensionality reduction processing;

[0024] Step 34: Remove edges with weights below the stability threshold ε from the graph structure to generate a sparse optimized structure graph;

[0025] Step 35: Input the node set V and the optimized edge set into the graph neural network for embedding learning to generate node embedding representations.

[0026] Firstly, in step 4, edge weights are computed using a graph attention network:

[0027]

[0028] Among them, w ij The edge weight between node i and its neighboring node j is represented by exp(), which is the natural exponential function e. x T is the transpose. Let be the original feature vector of node i. Let be the original eigenvector of node j, and W be the transformation matrix. is the attention weight, || denotes concatenation, N(i) is the neighbor set, and LeakyReLU is the linear rectified activation function with leakage.

[0029] Prioritizing step 6, an incremental learning mechanism is introduced into the health status knowledge graph to update the health status knowledge graph based on the multi-source status data collected at the current moment, including:

[0030] Based on the graph update triggering mechanism, if new feature variables not covered by the health status knowledge graph appear in the multi-source state data collected at the current moment, the new feature variables are used to generate corresponding new nodes, and new edges of the health status knowledge graph are constructed based on the correlation, collaboration or fault co-occurrence relationship between the new nodes and existing nodes.

[0031] A graph neural network is used to embed new nodes and their adjacent edges, and then the embeddings are integrated into the overall embedding representation of the original graph structure to obtain the node embedding results and the updated graph structure.

[0032] An edge weight decay mechanism is used to automatically reduce the weight of edges that have not been active for a long time. When the edge weight is lower than a preset threshold, an edge pruning operation is performed.

[0033] Prior to this, step 7 includes:

[0034] Based on the updated structural weights and node scores, calculate the transformer's overall health score at the current moment:

[0035]

[0036] Among them, S i S represents the health score value S of the i-th per unit value. i ; Numerator ∑ j∈N(i) w ij The denominator represents the structural importance of node i. Let w be the sum of edge weights of the entire graph, N(i) be the set of neighbors of node i, N(k) be the set of neighbors of node k, and w be the sum of edge weights of the entire graph. ij w represents the edge weight between node i and its neighboring node j. kj Let represent the edge weight between node k and its neighboring node j.

[0037] Prior to this, step 7 includes:

[0038] When the transformer's comprehensive health score is H total When the value is ≥90, the transformer is considered to be in healthy operating condition;

[0039] When 80≤H total When the value is less than 90, the transformer's operating status is determined to be a mild warning.

[0040] When 70≤H totalWhen the temperature is below 80°C, the transformer's operating status is determined to be a moderate warning level.

[0041] When 60≤H total When the value is less than 70, the transformer's operating status is determined to be a severe warning.

[0042] When H total When the value is less than 60, the transformer's operating status is determined to be abnormal.

[0043] Preferably, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described herein.

[0044] Preferably, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described herein.

[0045] The beneficial effects achieved by this invention are as follows:

[0046] This invention constructs an indicator system encompassing multiple physical characteristics such as electricity, heat, and vibration based on multi-source state data of transformers. It uses operational standards to determine health level judgment rules and early warning thresholds for various key indicators in the multi-source state data, forming a structured state knowledge graph. Then, a graph attention network is introduced to weight the edges connecting various indicators in the health state knowledge graph, fully exploring the causal dependencies between feature indicators under different operating states, achieving structural perception and causal modeling under complex operating conditions. This invention designs an edge-weighted node information update mechanism combined with an incremental learning strategy, enabling the evaluation system to adapt to the dynamic evolution of operating states. Based on the graph neural network embedding results and the transformer's comprehensive health score, this invention quantifies the current health state of the transformer and outputs a comprehensive health score. This method can comprehensively reflect the health change trend of transformers under various operating condition disturbances in actual operation, possessing strong engineering practicality, interpretability, and scalability, and is suitable for application scenarios such as transformer online monitoring systems and operation and maintenance decision support platforms. Attached Figure Description

[0047] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the present invention;

[0049] Figure 2 The diagram shows the structure of the attention network.

[0050] Figure 3 A specific scoring trend chart for the case. Detailed Implementation

[0051] See Figure 1 This application discloses an intelligent health status assessment method for power transformers based on knowledge graphs and graph neural networks. It is primarily applicable to the intelligent perception and health assessment of power transformer operating status, and exhibits good adaptability and versatility, especially in operating environments characterized by simultaneous acquisition of multi-source status data, significant data heterogeneity, and complex status changes. The method includes the following steps:

[0052] Step 1: Preprocess the pre-collected multi-source state data of the transformer operation to obtain per-unit values, and construct a standardized dataset based on the per-unit values. The multi-source state data of the transformer operation includes electrical parameters, thermal parameters, and mechanical parameters, where current represents the load condition, oil temperature represents the heat accumulation level, and vibration represents the mechanical stress state. Preprocessing includes imputing missing values, removing outliers, and normalizing the multi-source state data.

[0053] Step 2: Establish scoring rules based on standards and experience, convert the per-unit values ​​in the standardized dataset into health score values, and label the health score values ​​with health level labels, thereby obtaining a training dataset with label information based on the standardized dataset;

[0054] Step 3: Using the training dataset obtained in Step 2, construct a health status knowledge graph, with status indicators as nodes and physical or statistical associations as edges, representing the influence relationship between each feature; the graph structure of the health status knowledge graph is G = (V, E), where V is the set of nodes and E is the set of edges;

[0055] Step 4: Based on the training dataset obtained in Step 2 and the health status knowledge graph constructed in Step 3, the edge weights are calculated using a graph attention network;

[0056] Step 5: Based on edge weights, the node embedding representation is iteratively updated using a node feature aggregation mechanism guided by edge weights to obtain a node representation that integrates structural and feature information. This method enhances the correlation between different operational indicators, providing a more accurate feature foundation for subsequent health status assessment. The embedding representation generates a set of low-dimensional vectors for each node, used to comprehensively represent the node's feature attributes and its topological relationship in the graph structure. Based on the training dataset formed in Step 2 and the graph constructed in Step 3, this step uses the edge weights learned in Step 4 to perform node information aggregation and embedding updates. Specifically, for any node i, the feature vectors of each node in its neighborhood N(i) are weighted and summarized according to the corresponding edge weights, and then updated by linear transformation and nonlinear activation with its own features. The above process can be iterated multiple times to enhance the expressive ability of higher-order neighborhood dependencies. If necessary, multi-head attention, residual connections, and layer normalization can be introduced to improve training stability and expression upper limit. The resulting node embedding matrix simultaneously encodes the structural relationship and observation features between indicators, which will serve as the input for subsequent comprehensive health scoring and level determination.

[0057] Step 6: Based on the health status knowledge graph, to further enhance the model's ability to continuously adapt to changes in operating status, an incremental learning mechanism is introduced into the health status knowledge graph. This enables the health status assessment model to self-evolve as data is dynamically updated. During the continuous updating of multi-source status data, the graph structure and node representation are dynamically adjusted to adapt to the temporal evolution of equipment status. Specifically, this includes updating the health status knowledge graph based on the multi-source status data collected at the current moment, maintaining the stability of old state features while achieving gradual learning and memory transfer of new states. This mechanism aims to solve the problem that traditional models, after static training, cannot adapt to state transitions during long-term equipment operation, thus improving the model's online learning and generalization capabilities. Specifically, each node in the knowledge graph is considered to represent a physical quantity or key indicator in the transformer's operating status, and edges represent statistical or physical correlations between different indicators. As operating data is continuously input, the weights of these edges will dynamically evolve. To accurately characterize the temporal trend of correlation strength, the following edge weight update strategy is adopted:

[0058] Suppose that at time step t, the edge weight between node i and node j is When new data Δw ij Upon arrival, the edge weight increment Δw is first recalculated using the GAT model. ij The edge weight update formula is:

[0059]

[0060] Here, λ∈[0,1] is the time decay coefficient, which controls the balance between historical weights and contributions from new data. A larger λ value helps maintain the time stationarity of edge weights and suppresses the impact of occasional data disturbances; while a smaller λ value improves the responsiveness to new data patterns.

[0061] After the edge weights are updated, the system will automatically detect whether the graph structure has undergone topological changes. When a new input sample contains an indicator state that has not yet been involved in the historical graph, the indicator state will be used as a new node and an edge relationship will be established with the relevant node. For edge relationships that have not been activated for a long time, the system will reduce their edge weights through an exponential decay mechanism until they fall below a set threshold, triggering a structure simplification operation, thereby achieving automatic removal of redundant edges in the graph.

[0062] Based on the graph structure update results, this embodiment further incrementally adjusts the node embedding representation to support the feature input of the downstream health status prediction model. Considering the balance between computational efficiency and model real-time performance, the embedding vector update operation is only performed on the nodes involved in the structural changes and their local neighborhoods. Let the embedding vector of node i at time t be... Its update formula is:

[0063]

[0064] Where N(i) represents the set of neighbors of node i, α ij Let W be the attention coefficients after edge weight normalization, and W be the trainable weight matrix.

[0065] The aforementioned embedding update mechanism is triggered only in the following two situations: ① the neighborhood edge weights of a node change significantly; ② the Euclidean distance between the current embedding vector and the previous embedding vector exceeds a preset threshold δ. This design effectively avoids resource waste caused by frequent small fluctuations while ensuring model adaptability.

[0066] Step 7: Calculate the transformer's comprehensive health score H based on the weight coefficients determined by the node scores and edge weights. total :

[0067]

[0068] Where, α i S represents the node weight. i The score is given for the i-th indicator, and n is the total number of indicators.

[0069] The health status knowledge graph is a structure graph including multiple nodes and multiple edges. The nodes are used to represent the transformer's operating indicators, physical components, or health status labels. The edges are used to represent the relationships between the nodes. The relationships include, but are not limited to, causal relationships, cooperative relationships, and combinatorial relationships. Each edge is assigned an initial edge weight. The initial edge weight can be determined based on expert experience, correlation analysis of historical operating data, or fault statistics to reflect the connection strength between the nodes.

[0070] Graph Attention Network (GAT), an adaptive graph neural network method, is used in this invention to dynamically calculate edge weights in a health state knowledge graph. This method introduces an attention mechanism, allowing the influence strength between nodes to be adjusted in real time according to changes in operating conditions, thereby improving the graph structure's responsiveness to actual conditions. Specifically, GAT uses attention coefficients to quantify the interactions between nodes, enabling the weights of corresponding edges to be adaptively updated as the transformer's operating state changes. For example, under high load conditions, the influence of the "load rate" node on the "winding temperature" node is enhanced, and the model automatically increases the weight of the edge between them; while under low load conditions, the edge weight is correspondingly weakened, thus achieving dynamic adjustment of edge weights. The network structure of the graph attention network is as follows: Figure 2 As shown. First, the similarity coefficient between the target entity and each neighbor needs to be calculated. A shared weight matrix W is trained to increase the dimensionality of the input features and map them to the output features. The high-dimensional feature representations of the target node i and each neighbor node j after being mapped by the weight matrix W are concatenated. The concatenated high-dimensional feature vector is then mapped to a real value through a feedforward neural network, and the absolute attention coefficient is obtained by specifying an activation function.

[0071]

[0072] in, and Let W represent the embedding of the center node and the embedding of the neighbor node, respectively, and let W ∈ R × F′ be the shared weight matrix. This is the weight matrix. The edge weights are obtained by normalizing the neighborhood of the center node:

[0073]

[0074] Where N represents the set of all nodes adjacent to the center node i. The attention coefficients of all neighboring nodes are weighted and summed, and after activation, the output features of the center node i are obtained:

[0075]

[0076] Where σ(·) is the LeakyReLU activation function. This refers to the updated node features.

[0077] Step 1 includes: First, constructing a multi-dimensional transformer operating characteristic index library, which includes electrical parameters, thermal parameters, and mechanical parameters. Electrical parameters include load current and voltage fluctuations; thermal parameters include top oil temperature and ambient temperature; and mechanical parameters include core and tank vibration acceleration. Combining equipment operating standards and engineering experience, determine the reference range and optimal range for the electrical, thermal, and mechanical parameters. If an electrical, thermal, or mechanical parameter is outside the reference range, it is marked as abnormal. If the number of times an abnormal parameter is marked exceeds a preset threshold, the abnormal electrical, thermal, or mechanical parameters are removed and replaced with robust interpolation values. The electrical, thermal, and mechanical parameters are normalized to per-unit values ​​to obtain a standardized dataset.

[0078] Step 2 includes:

[0079] The health score S of the i-th item is calculated by performing standard scoring on each indicator according to the following scoring function. i :

[0080]

[0081] Among them, S i X represents the health score value of the i-th per-unit value. i x represents the measured value of the i-th per-unit value in the per-unit values. i,opt Let x represent the optimal operating value of the i-th per-unit value. i,min k represents the minimum acceptable value for a per-unit value. i This is the score decay coefficient, which controls the rate at which the score decreases.

[0082] Secondly, the health scores are categorized according to attributes, and weight factors are set to form a standardized indicator system, providing basic information for the subsequent construction of a structured health status knowledge graph.

[0083] Finally, based on pre-set labels and historical abnormal operation and maintenance data, health level labels are marked on the health score values ​​to obtain a training dataset with label information, thereby realizing the unified organization and label management of multi-source heterogeneous data of transformers. The historical abnormal operation and maintenance data mentioned in this application refers to abnormal state-related data collected and recorded during transformer operation due to faults, warnings, or operation and maintenance intervention events, specifically including but not limited to the following categories: ① Electrical parameters, such as three-phase current, voltage imbalance rate, and core grounding current, will exhibit abnormal fluctuations when overloaded or when insulation deteriorates, with the change amplitude exceeding the preset threshold; ② Thermal characteristic parameters, such as top oil temperature and winding temperature difference, may continuously exceed the warning threshold when cooling is insufficient or long-term overload occurs; ③ Mechanical and vibration parameters, such as core or tank vibration amplitude, will significantly increase when winding loosening or mechanical defects occur.

[0084] The training dataset generated by this method has complete indicators, clear scoring logic, and health level labels, making it suitable for embedded graph structure modeling, graph neural network training, and incremental state assessment scenarios. The graph neural network model has a multi-layer structure, and residual connections and activation functions can be combined between layers to enhance feature representation capabilities. The embedded calculation results can be concatenated or weighted after the outputs of each layer to further improve the discriminative ability of the final node representation.

[0085] Step 3 includes the following steps:

[0086] In constructing the edge set, expert knowledge and engineering experience rules are initially adopted, combined with physical dependencies and operational causal relationships, to pre-determine the connection relationships between some characteristic indicators. For example, there is a stable coupling between load current and oil temperature due to thermal effects, a mechanical-thermal linkage path exists between winding temperature difference and housing vibration, and a hysteresis response relationship may form between ambient temperature and top oil temperature. The connections established under these physical or empirical rules form the initial edge structure, providing a convergent search space for subsequent edge weight adjustment and pruning based on learning mechanisms.

[0087] Furthermore, when constructing the initial graph structure, it is necessary to avoid redundant propagation problems caused by over-connection and dense structure. Therefore, only node pairs with clear semantics or statistical significance are retained as members of the initial edge structure. This graph structure will serve as the basic graph structure input in subsequent steps, participating in edge weight learning, feature aggregation, and health status score propagation calculation in the graph neural network.

[0088] Step 31: Using health score as the node candidate set, analyze the features in the training dataset, calculate the lagged correlation, mutual information index and Granger causality coefficient of the features, and construct the correlation matrix between features;

[0089] Step 32: Based on the preset edge construction criteria, process the candidate nodes v in the node set V. iWith candidate node v j Perform a two-way test; if the correlation index of the candidate node is greater than the set threshold θ... r Furthermore, if the causality test is passed, then at candidate node v i With candidate node v j Constructing a directional edge e between them ij Obtain the optimized edge set and initialize the edge weights. As for the edge e ij Standardized relevant values;

[0090] Step 33: Extract the corresponding running feature vectors for all nodes in the graph structure G = (V, E). R d For a real space of dimension d, for the eigenvectors Perform normalization and dimensionality reduction processing;

[0091] Step 34: Remove edges with weights below the stability threshold ε from the graph structure to generate a sparse optimized structure graph, thereby improving the propagation efficiency of the subsequent graph neural network;

[0092] Step 35: Input the node set V and the optimized edge set into the graph neural network for embedding learning to generate node embedding representations containing semantic information and structural relationships, which serve as the input basis for subsequent health score calculation;

[0093] This method can effectively eliminate redundant index relationships and weakly correlated pathways, improve the discriminative ability and embedding expression quality of knowledge graph structures, and is suitable for causal mapping scenarios between large-scale monitoring indicators.

[0094] Step 4: Calculate edge weights using a graph attention network to represent the relative influence between nodes. The formula for calculating edge weights is:

[0095]

[0096] Among them, w ij The edge weight between node i and its neighboring node j is represented by exp(), which is the natural exponential function e. x T is the transpose. Let be the original feature vector of node i. Let be the original eigenvector of node j, and W be the transformation matrix. is the attention weight, || denotes concatenation, N(i) is the neighbor set, and LeakyReLU is the linear rectified activation function with leakage.

[0097] Step 6: Introduce an incremental learning mechanism into the health status knowledge graph, updating the health status knowledge graph based on the multi-source status data collected at the current moment, including:

[0098] Based on the graph update triggering mechanism, if new feature variables not covered by the health status knowledge graph appear in the multi-source status data collected at the current moment, the graph structure expansion is triggered: new nodes are generated using the new feature variables, and new edges of the health status knowledge graph are constructed based on the correlation, collaborative relationship or fault co-occurrence relationship between the new nodes and existing nodes. The initial edge weights of the new graph edges are set based on statistical correlation or expert experience.

[0099] A graph neural network is used to embed new nodes and their adjacent edges, and then integrates them into the overall embedding representation of the original graph structure to obtain the node embedding results and the updated graph structure, thereby achieving consistent updating of the graph structure.

[0100] To avoid excessive expansion of the graph structure, an edge weight decay mechanism is adopted to automatically reduce the weight of edges that have not been active for a long time. When the edge weight is lower than a preset threshold, an edge pruning operation is performed. A time decay factor for graph edge weights is set to gradually reduce the weight of edge relationships that have not been active for a long time, so as to reflect the decay trend of their importance in the current state.

[0101] Specifically, if an edge has not been updated within 50 consecutive sampling periods, it is considered to be inactive for a long time.

[0102] Based on the updated graph structure and node embedding results, the transformer comprehensive health score is calculated to reflect the comprehensive status of the latest operating conditions and to achieve adaptive updating of the status assessment results.

[0103] This method is easy to deploy, computationally efficient, and capable of continuous structural evolution, making it suitable for state assessment tasks in scenarios where transformer operating data is dynamically updated.

[0104] Based on the updated structural weights and node scores, calculate the transformer's overall health score at the current moment:

[0105]

[0106] Among them, H total This represents the transformer's overall health score at the current moment, with a value ranging from 0 to 100; S i S represents the health score value S of the i-th per unit value. i ; Numerator ∑ j∈N(i) w ij The denominator represents the structural importance of node i. Let w be the sum of edge weights of the entire graph, N(i) be the set of neighbors of node i, N(k) be the set of neighbors of node k, and w be the sum of edge weights of the entire graph. ij w represents the edge weight between node i and its neighboring node j. kjLet represent the edge weight between node k and its neighboring node j. The structure implements a scoring weighting strategy based on the degree of mutual influence of indicators; it tracks historical trends and analyzes fluctuations in the scores to determine the magnitude and rate of score changes, identify potential abnormal evolution trends, and trigger an early warning mechanism when score fluctuations exceed a set threshold. This comprehensive scoring model can quantitatively evaluate the health status of transformers under different operating conditions, achieving a health score driven by structural weights, and improving the evaluation model's ability to model the coupling relationships of multiple indicators and its physical interpretability.

[0107] The comprehensive health score of the transformer is mapped to five status levels according to the numerical range to achieve qualitative judgment of the current operating status of the power transformer. This includes: the score value is a continuous value limited to a preset numerical range, and the preset numerical range of the score range is 0 to 100, which is used to standardize the score results under different equipment or operating scenarios; the score range can be determined based on industry standards, operating experience or big data analysis results, and is adjustable.

[0108] When the transformer's comprehensive health score is H total When the value is ≥90, the transformer is considered to be in healthy operating condition;

[0109] When 80≤H total When the value is less than 90, the transformer's operating status is determined to be a mild warning.

[0110] When 70≤H total When the temperature is below 80°C, the transformer's operating status is determined to be a moderate warning level.

[0111] When 60≤H total When the value is less than 70, the transformer's operating status is determined to be a severe warning.

[0112] When H total When the value is less than 60, the transformer's operating status is determined to be abnormal.

[0113] By continuously monitoring and analyzing the trend of transformer health score changes, potential abnormal evolution processes can be identified, early warning signals can be issued in a timely manner, and a health assessment report can be provided. The health assessment report includes the current comprehensive health score of the transformer, the score level, the main influencing indicators, a summary of changes in the graph structure, and recommended operation and maintenance operations. The score levels include healthy, mild warning, moderate warning, severe warning, and abnormal state.

[0114] The system also supports writing each scoring result, along with original indicators, embedding vectors, edge weight distribution, and other information, into the health status archive, and synchronously updating the historical version of the knowledge graph, achieving traceability, comparability, and archiving of health assessment results. The assessment results are recorded and a long-term health archive is formed, including the historical sequence of scoring values, the evolution path of status levels, and a knowledge graph structure change log, which can be used for subsequent trend analysis and operation and maintenance strategy optimization. This archiving mechanism not only facilitates operation and maintenance personnel in reviewing past evolution processes but also provides historical data support for subsequent operation and maintenance decisions, further enhancing the application completeness and engineering practical value of this invention.

[0115] The method is applicable to various types of transformer equipment, including oil-immersed, dry-type, and combined types, and supports different voltage levels from 10kV to 500kV. It has good versatility and deployability, and can meet the actual application needs of various users such as power grid companies, power plants, and equipment manufacturers.

[0116] In this embodiment of the application, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0117] In this application embodiment, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0118] The computer equipment can be optionally integrated into the dispatch center server, substation main control terminal, edge processing equipment or embedded device, and can adapt to various operating modes such as distributed deployment, centralized evaluation and remote diagnosis, and has good engineering implementation flexibility.

[0119] Example 2

[0120] To verify the effectiveness and adaptability of the transformer health status assessment method proposed in this invention in practical scenarios, online monitoring data of multiple physical quantities from an ODES-334000 / 500 transformer in Shanxi Province from November to December 2024 were selected as the research object. The data included multiple source status data such as current, voltage, on-load tap changer, winding temperature, oil temperature, gas content, and vibration amplitude. These multi-source status data broadly represent the overall characteristics of the equipment's operating status under typical physical fields such as heat, vibration, and electrical insulation. Table 1 lists the initial reference values ​​and warning thresholds of some key features of the multi-source status data. The experimental period covered 60 days, with oil temperature and partial discharge data sampling period set at once every 1 hour, and oil gas sampling rate at once every 12 hours.

[0121] Table 1. Reference Values ​​and Early Warning Thresholds for Key Transformer Operation Indicators

[0122]

[0123]

[0124] The threshold values ​​for all the above indicators are set based on authoritative standards and long-term engineering application experience, possessing clear technical basis and operability. Specifically, the criteria for determining the concentration of combustible gases in the oil refer to the "IEEE C57.104 Guideline for Dissolved Gas Analysis in Transformer Oil," which systematically classifies the concentration levels of different types of gases and their corresponding anomaly levels. Methane (CH4) is set at 100 μL / L to identify early low-energy fault signals; ethane (C2H6) has a limit of 35 μL / L, serving as an auxiliary criterion for judging thermal faults; hydrogen (H2) has a threshold of 300 μL / L, primarily used to detect latent risks such as high-temperature overheating and partial discharge; acetylene (C2H2), as an important indicator for judging arc faults, has a threshold of 10 μL / L, the presence of which indicates that the equipment may be in a relatively serious abnormal state.

[0125] The upper limit for the top oil temperature warning is set at 85℃, referring to the oil temperature control requirements in "DL / T 572-2010 Operating Procedures for Power Transformers" and determined in conjunction with the thermal stability capability of the equipment under actual operating conditions; the load rate warning threshold is 100%, which is the upper limit of the rated capacity, limited according to "GB / T 1094.1-2013 Power Transformers Part 1: General Rules", exceeding this value is considered overload operation; the three-phase unbalance rate limit is 4%, referring to "GB / T 15543-2008 Three-phase voltage unbalance", used to constrain the local thermal effects that may be caused by current asymmetry; the phase difference of the winding DC resistance is controlled within 3%, according to "DL / T 596-1996 Preventive Test Procedures for Power Equipment", used to judge the internal consistency of the winding and process deviation.

[0126] The warning value for core grounding current is 0.3A, set based on typical substation operation logs, used to identify core insulation degradation and abnormal grounding; the warning value for oil breakdown voltage is 35kV, according to GB / T507-2002 "Test Method for Breakdown Voltage of Insulating Oil", below this value is considered a decline in oil insulation capacity; the load current threshold is set to 1600A based on the rated operating capacity of the transformer model, exceeding which will have a significant impact on the thermal field and electrical stress; the phase-to-phase difference in winding resistance is set to no more than 3% to ensure winding process consistency and symmetry, and prevent localized heating or abnormal voltage distribution; the breakdown voltage of transformer oil during operation is set to 35kV, based on typical substation operation logs, if the detected value is below this threshold, it indicates that the insulation system is at risk of degradation, and maintenance or replacement should be carried out in a timely manner.

[0127] All of the above thresholds comply with the technical specifications of relevant national or international standards, meet the actual engineering needs of power transformer operation status monitoring, safety early warning and fault prevention, and provide reliable physical constraint boundaries and judgment basis for the health assessment model proposed in this invention.

[0128] The period from November 20th to December 10th, 2024, was selected as the typical analysis time window. Combining equipment operation and maintenance records and online monitoring data during this period, a trend retrospective analysis of the health score results was conducted. Through systematic analysis of the dynamic evolution of the health score curve during this period, significant segmented fluctuations were observed before and after the introduction of typical event nodes. Based on the score change trend, it can be divided into three stages, as detailed below. Figure 3 As shown, this reflects the different changes in the operating conditions of the transformer.

[0129] (1) From November 24 to November 29, 2024, the environmental conditions in this region underwent significant seasonal changes. Specifically, the average temperature gradually decreased from 12℃ to 0℃, the average humidity decreased from 58% to 45%, the wind force increased from level 1 to level 4, and the sunshine duration increased from 6.0 hours to 7.0 hours. During the same period, the transformer load rate remained at a low level, slowly decreasing from 28% to 22%, thus essentially ruling out the possibility of load fluctuations affecting the equipment's operating status. Given the stable load, the continuous rise in the transformer health score during this period has raised concerns and requires detailed analysis from an environmental perspective. Relevant meteorological data obtained from the China Meteorological Administration website are shown in Table 2:

[0130] Table 2. Statistical Table of Relevant Meteorological Data

[0131]

[0132] First, the gradual decrease in ambient temperature improves the heat exchange conditions between the transformer and the environment, significantly enhancing cooling efficiency. This leads to a continuous decrease in the temperature of the insulating oil and windings, effectively slowing down the insulation aging process and reducing the rate of combustible gas generation in the oil, thus promoting the stability of dissolved gas analysis indicators. Second, the decrease in relative humidity reduces the probability of condensation on equipment surfaces, optimizes the surface condition of the insulation system, reduces the risk of partial discharge, and is conducive to the continuous optimization of health status. In addition, increased wind speed enhances the natural convection heat dissipation effect, further reducing the temperature rise of the equipment, especially under medium and low load conditions, which is reflected in the steady increase of the score value. Changes in sunshine duration also buffer the diurnal temperature difference, reducing the fluctuation range of thermal parameters, thereby improving the stability of the operating status. It can be seen that the upward trend of the score in this stage is mainly driven by the comprehensive improvement of multiple environmental factors such as temperature, humidity, wind speed, and sunshine. It is characterized by a continuous, gradual, and non-abrupt upward trend in the score, reflecting the responsiveness and physical consistency of the scoring model to changes in the external environment.

[0133] (2) From December 1st to December 6th, 2024, the transformer health score showed a continuous downward trend, decreasing from 96.2 to 91.1. The curve changed smoothly overall, without any sudden jumps. According to the operation and maintenance log records during the same period, starting from December 1st, due to seasonal load fluctuations, the low-voltage side reactive power compensation device began to switch on and off at a higher frequency to maintain the stability of the system voltage and power factor. Although this type of reactive power adjustment behavior complies with the operating specifications, frequent switching operations can easily cause short-term voltage disturbances and transient current fluctuations, which in turn cause slight disturbances to the electromagnetic stress of the transformer windings, the electrical insulation system, and the thermal field distribution. Although the changes in relevant parameters did not exceed the technical allowable limits and did not trigger the safety alarm mechanism, such disturbances were identified as a sign of slight deterioration in the operating status in the health scoring system, manifested as a slow decline in the score.

[0134] Furthermore, a brief rebound was observed in the scoring data for December 4th. Analysis of the day's operation logs revealed a significant decrease in the frequency of reactive power compensation switching during this period, indicating that the equipment's operating status had returned to a relatively stable state, and the fluctuation range of thermoelectric parameters had decreased, leading to a slight increase in the health score. These phenomena demonstrate that the scoring model proposed in this invention can effectively perceive the impact of frequent adjustment behaviors on the equipment's operating status and possesses a sensitive response capability to short-term changes in operational stability, exhibiting high dynamic adaptability and physical consistency.

[0135] (3) Starting from December 7, 2024, the transformer health score entered a period of continuous recovery. Based on the analysis of the operation and maintenance records during this period, the operation strategy of the low-voltage side reactive power compensation device was adjusted from the previous high-frequency switching mode to a conventional on-demand compensation mode. The frequency of device switching on and off significantly decreased, and key operating parameters of the electrical system, such as voltage and current, gradually stabilized. Affected by this adjustment measure, the overall operating status of the transformer gradually returned to steady-state conditions, and various representative physical indicators gradually returned to the normal operating range. The health score also showed a stable upward trend.

[0136] Compared to the continuous decline in scores in the previous stage, the recovery process in this stage exhibits a certain lag and continuity, consistent with the physical recovery mechanism characteristics of electrical equipment after experiencing disturbances. Particularly in the early stages of recovery, the score increase is relatively gradual, indicating that the constructed health scoring model fully considers the operational inertia and state evolution laws of the equipment in its design, avoiding over-response of the scoring system due to instantaneous improvement. This characteristic demonstrates the good balance achieved by the model proposed in this invention between dynamic response and scoring stability, possessing strong practical adaptability and engineering application value.

[0137] In summary, through the comparative analysis of the scoring trends and operation and maintenance records in the above three stages, it can be fully demonstrated that the health status scoring model constructed in this invention can not only accurately perceive changes in transformer status, but also maintain reasonable sensitivity and stability under slight disturbance conditions, and has good engineering usability and consistency with physical interpretation.

[0138] To further verify the performance advantages of the health status assessment method proposed in this invention under complex working conditions, two traditional assessment methods are introduced as comparison benchmarks, including: (1) a hierarchical judgment method based on expert rules and fixed thresholds; and (2) a fuzzy comprehensive evaluation method constructed using fuzzy membership degree and weighted average. Both of these methods are commonly used status assessment methods in current engineering sites, representing static rule-driven and semi-empirical scoring models respectively, and are therefore representative.

[0139] While maintaining consistency in monitoring data sources, indicator systems, and sampling periods, the model of this invention and two traditional methods were applied to perform health scoring and anomaly identification on the target transformer's operating status. Four quantitative performance indicators were examined: (1) accuracy of anomaly identification; (2) response delay time of the scoring curve to abnormal disturbances; (3) standard deviation of the scoring output; and (4) goodness of fit R between the overall scoring trend and the operation and maintenance event log. 2 The experimental comparison results are as follows:

[0140] Table 3. Statistical Table of Relevant Meteorological Data

[0141]

[0142] The various embodiments in this specification are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0143] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention described herein. This application is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not invented herein. The specification and embodiments are to be considered exemplary only.

[0144] The above specific embodiments further illustrate the purpose, technical solution and beneficial effects of this application. It should be understood that the above are only specific embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solution of this application should be included within the scope of protection of this application.

Claims

1. A method for intelligent assessment of the health status of power transformers based on knowledge graphs and graph neural networks, characterized in that, include: Step 1: Preprocess the multi-source status data of the transformer operation collected in advance to obtain per-unit values, and construct a standardized dataset based on the per-unit values; Step 2: Convert the per-unit values ​​in the standardized dataset into health score values, label the health score values ​​with health level labels, and obtain the training dataset; Step 3: Using the training dataset obtained in Step 2, construct a health status knowledge graph, with status indicators as nodes and physical or statistical associations as edges; the graph structure of the health status knowledge graph is G = (V, E), where V is the set of nodes and E is the set of edges; Step 4: Based on the training dataset obtained in Step 2 and the health status knowledge graph constructed in Step 3, the edge weights are calculated using a graph attention network; Step 5: Based on edge weights, use the edge weight-guided node feature aggregation mechanism to iteratively update the node embedding representation to obtain a node representation that integrates structural information and feature information; Step 6: Introduce an incremental learning mechanism to the health status knowledge graph, and update the health status knowledge graph based on the multi-source status data collected at the current moment.

2. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 1, characterized in that, Step 1 includes: A multi-dimensional transformer operation characteristic index library is constructed, which includes electrical parameters, thermal parameters, and mechanical parameters. Electrical parameters include load current and voltage fluctuations; thermal parameters include top oil temperature and ambient temperature; and mechanical parameters include core and tank vibration acceleration. Reference and optimal ranges for electrical, thermal, and mechanical parameters are determined. If an electrical, thermal, or mechanical parameter is outside the reference range, it is marked as abnormal. If the number of times an abnormal parameter is marked exceeds a preset threshold, the abnormal electrical, thermal, or mechanical parameter is removed and replaced with a robust interpolation value. The electrical, thermal, and mechanical parameters are normalized to per-unit values ​​to obtain a standardized dataset.

3. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 1, characterized in that, Step 2 includes: Calculate the health score S of the i-th per-unit value. i : Among them, S i X represents the health score value of the i-th per-unit value. i x represents the measured value of the i-th per-unit value in the per-unit values. i,opt Let x represent the optimal operating value of the i-th per-unit value. i,min k represents the minimum acceptable value for a per-unit value. i This is the score decay coefficient; Based on attributes, health scores are categorized directionally, and weighting factors are assigned. Based on pre-defined labels and historical data of abnormal operation and maintenance, health level labels are labeled for health score values ​​to obtain a training dataset.

4. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 1, characterized in that, Step 3 includes the following steps: Step 31: Using health score as the node candidate set, analyze the features in the training dataset, calculate the lagged correlation, mutual information index and Granger causality coefficient of the features, and construct the correlation matrix between features; Step 32: For candidate nodes v in the node set V i With candidate node v j Perform a two-way test; if the correlation index of the candidate node is greater than the set threshold θ... r Furthermore, if the causality test is passed, then at candidate node v i With candidate node v j Constructing a directional edge e between them ij Obtain the optimized edge set and initialize the edge weights. As for the edge e ij Standardized relevant values; Step 33: Extract the corresponding running feature vectors for all nodes in the graph structure G = (V, E). Rd is a real space of dimension d, and for the feature vector Perform normalization and dimensionality reduction processing; Step 34: Remove edges with weights below the stability threshold ε from the graph structure to generate a sparse optimized structure graph; Step 35: Input the node set V and the optimized edge set into the graph neural network for embedding learning to generate node embedding representations.

5. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 1, characterized in that, Step 4: Calculate edge weights using a graph attention network. Among them, w ij The edge weight between node i and its neighboring node j is represented by exp(), which is the natural exponential function e. x T is the transpose. Let be the original feature vector of node i. Let be the original eigenvector of node j, and W be the transformation matrix. is the attention weight, || denotes concatenation, N(i) is the neighbor set, and LeakyReLU is the linear rectified activation function with leakage.

6. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 1, characterized in that, Step 6: Introduce an incremental learning mechanism into the health status knowledge graph, updating the health status knowledge graph based on the multi-source status data collected at the current moment, including: Based on the graph update triggering mechanism, if new feature variables not covered by the health status knowledge graph appear in the multi-source state data collected at the current moment, the new feature variables are used to generate corresponding new nodes, and new edges of the health status knowledge graph are constructed based on the correlation, collaboration or fault co-occurrence relationship between the new nodes and existing nodes. A graph neural network is used to embed new nodes and their adjacent edges, and then the embeddings are integrated into the overall embedding representation of the original graph structure to obtain the node embedding results and the updated graph structure. An edge weight decay mechanism is used to automatically reduce the weight of edges that have not been active for a long time. When the edge weight is lower than a preset threshold, an edge pruning operation is performed.

7. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 6, characterized in that, Step 7 includes: Based on the updated structural weights and node scores, calculate the transformer's overall health score at the current moment: Among them, S i S represents the health score value S of the i-th per unit value. i ; Numerator term ∑j ∈ N ( i ) wij represents the structural importance of node i, and the denominator is... Let w be the sum of edge weights of the entire graph, N(i) be the set of neighbors of node i, N(k) be the set of neighbors of node k, and w be the sum of edge weights of the entire graph. ij wkj represents the edge weight between node i and its neighboring node j, and wkj represents the edge weight between node k and its neighboring node j.

8. The intelligent assessment method for the health status of power transformers based on knowledge graphs and graph neural networks according to claim 7, characterized in that, Step 7 includes: When the transformer's comprehensive health score is H total When the value is ≥90, the transformer is considered to be in healthy operating condition; When 80≤H total When the value is less than 90, the transformer's operating status is determined to be a mild warning. When 70≤H total When the temperature is below 80°C, the transformer's operating status is determined to be a moderate warning level. When 60≤H total When the value is less than 70, the transformer's operating status is determined to be a severe warning. When H total When the value is less than 60, the transformer's operating status is determined to be abnormal.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 8.

Images