Model and nameplate information fused low-voltage distribution network data system analysis method

Abstract

The invention discloses a model and nameplate information fused low-voltage power distribution network data system analysis method, which comprises the following steps: acquiring nameplate parameters and operation state data of each low-voltage power distribution device, and constructing a multi-modal knowledge graph comprising device nodes, nameplate parameter nodes and operation data nodes; fusing the static nameplate features and the dynamic operation features through a knowledge graph neural network to generate a comprehensive representation vector; utilizing a cross feature modeling mechanism to extract a structural dependency relationship between nameplate fields, and generating a nameplate cross feature tensor; fusing the two types of features, inputting the features into a state analysis model, and predicting an overload risk score, an aging degree estimated value and an alarm priority level; and an analysis result is written back into the graph structure to form a continuously-updated structured operation state analysis record. According to the invention, the high-precision evaluation of the equipment operation state and the dynamic maintenance of the map structure can be realized, and the intelligent analysis and early warning capability of the low-voltage power distribution system is improved.

Description

Technical Field

[0001] This invention relates to the intersection of power information technology and artificial intelligence technology, and in particular to a method for analyzing low-voltage distribution network data systems by fusing model and nameplate information. Background Technology

[0002] As a crucial component of the distribution network's end point, the operating status of low-voltage power distribution equipment directly impacts the network's security and power supply stability. While nameplate parameters such as rated voltage, rated current, frequency, and manufacturing date reflect the equipment's factory attributes and design capability limits, they cannot dynamically depict the equipment's operating status and aging degree at different service stages. Consequently, traditional condition assessment methods based on nameplate parameters fail to reflect real-time operating characteristics.

[0003] Currently, the assessment of distribution network operation status is mostly based on monitoring data such as current, voltage, and temperature collected by SCADA systems or smart sensing terminals, and anomaly identification is performed through simple threshold judgments or rule engines. However, such methods lack the ability to model individual differences of equipment and struggle to understand the contextual semantics of equipment operation by combining historical nameplate information, resulting in poor generalization of judgment results and insufficient risk identification capabilities.

[0004] While existing research attempts to introduce data-driven algorithms or machine learning models for status identification, it often treats operational data and nameplate information separately, failing to establish a structured relationship between the two. This makes it impossible to uncover potential abnormal signs of equipment from the interaction dependencies between nameplate parameters and operational data. Furthermore, equipment information management systems and analysis algorithms are mostly deployed separately, lacking a unified information carrying and updating mechanism. This makes it difficult to achieve write-back and closed-loop updates of operational status assessment results, hindering the formation of a knowledge system for power distribution equipment oriented towards a closed-loop operation and maintenance system.

[0005] Therefore, how to provide a low-voltage distribution network data system analysis method that integrates model and nameplate information is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a data system analysis method for low-voltage distribution networks that integrates model and nameplate information. This invention constructs a multimodal knowledge graph comprising equipment nodes, nameplate parameter nodes, and operational data nodes, fusing static nameplate structural features with dynamic operational sequence features to improve the comprehensiveness and accuracy of equipment operational status analysis. By combining graph neural networks and attention mechanisms, it achieves deep modeling of the cross-relationships of nameplate parameters and the trends in operational status changes, outputting key indicators such as overload risk scores, aging degree estimates, and alarm priority levels. Through graph back-writing of prediction results, a closed-loop feedback mechanism for equipment state evolution is established, effectively enhancing the intelligent perception and risk warning capabilities of power distribution equipment. This invention can significantly improve the precision of equipment status identification, analysis efficiency, and management intelligence level in low-voltage power distribution systems.

[0007] A method for analyzing low-voltage distribution network data systems by fusing model and nameplate information according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect the nameplate parameters of each low-voltage power distribution equipment, organize the nameplate parameters into structured static fields, and record the operation data during operation in chronological order as a multi-dimensional dynamic operation sequence.

[0009] S2. Use the structured static fields of each low-voltage power distribution equipment and nameplate parameters, as well as the multidimensional dynamic runtime sequence of the operating data, as different types of nodes in the multimodal knowledge graph to construct a multimodal knowledge graph;

[0010] S3. Use a multimodal knowledge graph neural network to jointly learn the node information in the multimodal knowledge graph, perform feature aggregation processing on the device nodes through the multimodal knowledge graph neural network, and deeply fuse the static structural features in the nameplate parameters with the temporal change features in the operating data to generate a comprehensive representation vector;

[0011] S4. Extract four sets of static field pairs from the nameplate parameters as cross-feature groups. Through the cross-feature modeling structure of the multi-head attention mechanism, calculate the nonlinear relationship between the four sets of field pairs, extract the structural dependency features between the nameplate parameters, and output the cross-feature tensor of the nameplate parameters.

[0012] S5. Concatenate the comprehensive representation vector and the cross-feature tensor, and input them into the equipment operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Classify or regress the operation status, and output the overload risk score, aging degree estimate and alarm priority level.

[0013] S6. Synchronously write the prediction results back to the corresponding device nodes in the multimodal knowledge graph and update the multimodal knowledge graph structure.

[0014] Optionally, S1 specifically includes:

[0015] S11. Read the identification information on the external nameplate of each low-voltage power distribution equipment and collect six parameters: rated voltage, rated current, rated frequency, manufacturing date, installation location and insulation class.

[0016] S12. Use information acquisition tools to parse the nameplate parameters, convert the rated voltage, rated current and rated frequency values ​​into a standard format in volts, amperes and hertz, and organize the field names, numerical units and representation methods into a unified template structure and record it as a static structured field.

[0017] S13. Calculate the service life of the current equipment based on the manufacturing date field, and add the service life value as the initial aging parameter to the static structured field;

[0018] S14. Encode the installation location field to convert the descriptive geographic information into the substation, line zone, or area number where the low-voltage power distribution equipment is located.

[0019] S15. Connect to the operation monitoring terminal of low-voltage power distribution equipment to collect real-time operation data on current, voltage, power factor and temperature changes during operation;

[0020] S16. Add a corresponding timestamp to each piece of running data, and arrange the collected running data in chronological order to form a multi-dimensional dynamic running sequence.

[0021] Optionally, S2 specifically includes:

[0022] S21. Use the unique number of each low-voltage power distribution device as the central node identifier to create the corresponding device node;

[0023] S22. Based on the nameplate parameters bound in the equipment nodes, create six types of nameplate parameter nodes: rated voltage, rated current, rated frequency, manufacturing date, installation location, and insulation class. Connect each parameter node to the corresponding equipment node through attribute connection edges to form a one-to-many static association structure between low-voltage power distribution equipment and nameplate parameters.

[0024] S23. Divide the collected operating data of current, voltage, power factor and temperature changes into continuous time segments with equal time intervals in chronological order. Create an operating data node for each time segment and record the dynamic status value within the time window.

[0025] S24. Connect each running data node to the corresponding device node in sequence through time relationship edges, and mark the timestamp information to form a time-series dynamic connection relationship between the device and the running status.

[0026] S25. Integrate equipment nodes, nameplate parameter nodes, and operation data nodes into a unified multimodal knowledge graph.

[0027] Optionally, S3 specifically includes:

[0028] S31. Encode the text content and numerical fields of each device node, nameplate parameter node and operation data node in the multimodal knowledge graph. Convert the categorical fields into one-hot vectors. Subtract the preset minimum value from the rated voltage, rated current, rated frequency and manufacturing date fields respectively, and then divide by the difference between the maximum value and the minimum value to obtain normalized values ​​with a value range between 0 and 1. Finally, concatenate all categorical fields into a normalized node feature vector of the same length.

[0029] S32. In each device node, read all connected nameplate parameter nodes according to the pointing relationship of attribute edges. According to the type of nameplate parameter fields, map discrete fields to fixed-length binary vectors through one-hot encoding. Scale the values ​​of continuous fields to the [0,1] range through the max-min normalization method. Concatenate the encoding results in a preset order to generate normalized node feature vectors. Align all normalized feature vectors one by one in the dimension direction according to the element position. Add all the values ​​at the corresponding positions to obtain the summary vector. Divide the summary result by the number of nameplate parameter nodes to obtain the average vector and generate a static structural feature representation.

[0030] S33. In each device node, read the connected operating data nodes in chronological order, extract four operating parameters from each operating data node: current, voltage, power factor and temperature change. Use the rated current, rated voltage, rated frequency and historical maximum temperature rise value in the nameplate parameters as normalization reference values, and perform division normalization on the current collected values ​​to obtain the four normalized values.

[0031] S34. Arrange the four normalized values ​​of all nodes in chronological order as one record per row, and arrange them row by row to form a two-dimensional time series matrix with time as the row and index as the column, to generate a time series change feature matrix representing the running status.

[0032] S35. Input the static structural feature vector and the temporal change feature matrix into the multimodal knowledge graph neural network, and perform vector dot product operation with the feature representation vector of each low-voltage power distribution equipment respectively, and calculate the matching score between the static structural features and the temporal change features.

[0033] S36. Input the matching scores of static structural features and temporal change features into the Softmax function for normalization to obtain the attention weight coefficients of static structural features and temporal change features. Perform weighted summation on the two types of features according to the corresponding attention weight coefficients to generate the fused comprehensive representation vector.

[0034] Optionally, S4 specifically includes:

[0035] S41. Read the rated voltage field and manufacturing date field respectively. Obtain the maximum and minimum values ​​of the rated voltage field and manufacturing date field in all low-voltage power distribution equipment. According to the linear mapping formula between the maximum and minimum values, convert the field values ​​into normalized values. Then, concatenate the two normalized values ​​into a two-dimensional feature vector.

[0036] S42. Read the rated current field and the installation location field. Use the same normalization formula to convert the rated current field into a normalized value. Perform one-hot encoding on the installation location field to generate a one-hot encoded vector with a length equal to the number of location categories. Then concatenate the two vectors in order to form a two-dimensional feature vector.

[0037] S43. Normalize the rated voltage field and rated current field respectively and splice them to form a two-dimensional vector. Normalize the manufacturing date field and splice the insulation class field after thermal encoding to form another set of two-dimensional vectors.

[0038] S44. Input the obtained two-dimensional vector into multiple subheads in the multi-head attention mechanism. In each subhead, split the two-dimensional vector into two one-dimensional vectors. Multiply the corresponding values ​​of each dimension of the two one-dimensional vectors to obtain an intermediate vector. Sum all elements of the intermediate vector and use it as the original relevance score of the field pair in the current subhead.

[0039] S45. The scores of all field pairs in the same subhead are uniformly fed into the Softmax function for normalization to obtain the attention weights of the field pairs under the subhead. The attention weights are multiplied by each dimension of the original feature vector to obtain the weighted feature representation. The weighted feature vectors output by all subheads are concatenated and fed into a linear transformation network for dimension unification. Finally, the cross feature representation vectors of the field pairs are output. The above operation is repeated for the four sets of field pairs in turn. The four sets of output feature vectors together form the cross feature tensor of the nameplate parameters.

[0040] Optionally, S5 specifically includes:

[0041] S51. Arrange the comprehensive representation vector and the corresponding cross feature tensor according to their last dimension. Perform concatenation operation on each row of the two vectors according to the order of their vector dimensions. Arrange all the feature values ​​in the comprehensive representation vector by column. Then, concatenate all the feature values ​​in the cross feature tensor in the same order. After concatenation, generate a fused feature vector with a length equal to the sum of the dimensions of the original two sets of vectors.

[0042] S52. Input the fused feature vector into the input layer of the device operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Use a linear transformation function to perform feature mapping operation on the fused feature vector. Perform matrix multiplication on the input fused feature vector and the set weight matrix. Add a bias vector with the same dimension as the output dimension, which is automatically optimized together with the weight matrix during training. Output a fixed-length hidden layer feature representation to obtain the hidden feature vector.

[0043] S53. Apply the ReLU activation function to the hidden feature vector and perform threshold transformation element by element. Then, input the activated hidden feature vector into a multi-layer fully connected structure. Perform matrix multiplication and bias summation operations once in each layer and connect it to the ReLU activation function. Continuously reconstruct the combination relationship of features in each layer and output the feature result vector.

[0044] S54 inputs the feature result vector into three independent output branch structures. In the first branch, the input feature vector is multiplied and summed with a 128×1 weight vector, and then a scalar bias is added to generate a single value as an overload risk score. In the second branch, two linear mapping operations are performed in sequence. The first operation maps the feature result vector to an intermediate dimension vector to extract feature information related to the aging trend. The second operation compresses the intermediate vector into a single output value as an estimate of the degree of aging.

[0045] S55. In the third branch, after multiplying and biasing the input feature vector with the multidimensional classification weight matrix, the result vector is normalized using the Softmax function. The probability value of each alarm level is output, and the category corresponding to the highest probability is selected as the alarm priority level. The three types of operation status analysis results are output.

[0046] S55. For each device whose fusion feature vector input is the fusion of multimodal graph features and multidimensional nameplate interaction features, read the overload risk score, aging degree estimate, and alarm priority level, and concatenate the three analysis results into a structured output vector. Extract the device number field from the information table into the fusion feature vector input of the fusion multimodal graph features and multidimensional nameplate interaction features, and bind it one-to-one with the output vector to generate a structured analysis record. Sort the structured analysis records of all devices according to the data acquisition timestamp, write the sorted records line by line into the analysis report file, and output the operation status analysis report.

[0047] Optionally, S6 specifically includes:

[0048] S61. Establish a key-value correspondence between the unique number of the device node and the overload risk score, aging degree estimate and alarm priority level in the analysis results, and generate a result update dictionary;

[0049] S62. Locate the device node that matches the device number in the multimodal knowledge graph and read the attribute fields under the node;

[0050] S63. Write the three analysis results in the result update dictionary to the corresponding device node in the form of attribute key values, and update the overload score field, aging status field and alarm level field respectively.

[0051] S64. Mark the updated device node as "analyzed", read the time identifier corresponding to the current analysis batch, create an edge structure of type "analysis result attribution", point the edge from the device node to the batch node, and write the analysis timestamp in the edge attribute to record the analysis timestamp.

[0052] S65. Save the updated multimodal knowledge graph structure and output the updated multimodal knowledge graph structure.

[0053] The beneficial effects of this invention are:

[0054] This invention proposes a low-voltage distribution network data system analysis method that integrates model and nameplate information. Addressing existing problems in distribution equipment status analysis, such as low utilization of nameplate parameters, isolated operational data features, insufficient status identification accuracy, and inability to write back prediction results, this invention constructs a multimodal knowledge graph that integrates equipment nodes, nameplate parameter nodes, and operational data nodes. This graph unifies the representation of static structural information and dynamic operating status. A graph neural network is used for comprehensive feature extraction, and an attention mechanism is employed to uncover cross-dependencies between nameplate fields, improving the accuracy of aging and abnormal signal identification. Furthermore, the analysis results are synchronously written into the multimodal knowledge graph, enabling structured tracking of state evolution. This invention boasts advantages such as high information fusion depth, high precision in status identification, structured representation of prediction results, and synchronous graph updates. It effectively solves the problems of diagnostic lag and management inefficiency caused by data fragmentation and lack of closed-loop processing in low-voltage distribution equipment status analysis. Attached Figure Description

[0055] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0056] Figure 1 The flowchart shows a low-voltage distribution network data system analysis method that integrates model and nameplate information proposed in this invention.

[0057] Figure 2 This is a flowchart illustrating the construction process of a multimodal knowledge graph that integrates static nameplate parameters and dynamic operating status, as proposed in this invention.

[0058] Figure 3 The flowchart illustrates the structure update process for synchronously writing the operational status analysis results into the multimodal knowledge graph, as proposed in this invention. Detailed Implementation

[0059] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0060] refer to Figure 1-3 A method for analyzing low-voltage distribution network data systems by fusing model and nameplate information includes the following steps:

[0061] S1. Collect the nameplate parameters of each low-voltage power distribution equipment, organize the nameplate parameters such as rated voltage, rated current, rated frequency, manufacturing date, and installation location into structured static fields, and record the operating data of current, voltage, power factor, and temperature changes during operation into a multi-dimensional dynamic operating sequence according to the time order.

[0062] S2. Take the structured static fields of each low-voltage power distribution equipment and nameplate parameters, as well as the multidimensional dynamic runtime sequence of the operating data, as different types of nodes in the multimodal knowledge graph. Construct a multimodal knowledge graph containing equipment nodes, nameplate parameter nodes, and operating data nodes. Nameplate parameters are connected to equipment nodes through attribute edges, and operating data is connected to equipment nodes through time relationship edges, thus establishing a complete equipment-nameplate-state triple relationship structure.

[0063] S3. Use a multimodal knowledge graph neural network to jointly learn the node information in the multimodal knowledge graph, and perform feature aggregation processing on the device nodes through the multimodal knowledge graph neural network. Deeply fuse the static structural features in the nameplate parameters with the temporal change features in the operating data to generate a comprehensive representation vector. The representation vector can simultaneously express the device's own attributes, the nameplate capability boundary, and the real-time operating status.

[0064] S4. Extract the following four sets of static field pairs from the nameplate parameters as cross-feature groups: rated voltage and manufacturing date, rated current and installation location, rated voltage and rated current, and manufacturing date and insulation class. Take these four sets of two-dimensional feature pairs as input, calculate the nonlinear relationship between the four sets of field pairs through the cross-feature modeling structure of the multi-head attention mechanism, extract the structural dependency features between the nameplate parameters, and output the cross-feature tensor of the nameplate parameters.

[0065] S5. Concatenate the comprehensive representation vector and the cross-feature tensor, and input them into the equipment operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Classify or regress the operation status, and output the overload risk score, aging degree estimate and alarm priority level.

[0066] S6. Synchronously write the prediction results back to the corresponding device nodes in the multimodal knowledge graph and update the multimodal knowledge graph structure.

[0067] The system collects nameplate parameters and operational sequence data for each low-voltage power distribution device, constructs structured static fields and multidimensional dynamic sequences, and achieves unified modeling of basic equipment information and real-time status, improving data integrity. It constructs a multimodal knowledge graph from equipment nodes, nameplate parameter nodes, and operational data nodes, forming a triple structure of equipment-nameplate-status, enhancing information organization and correlation. A multimodal knowledge graph neural network is used to fuse static structural features and temporal change features, generating a powerful comprehensive representation vector to improve the accuracy of status modeling. Nonlinear interaction relationships between field pairs in the nameplate parameters are extracted to generate cross-feature tensors, strengthening the ability to model deep dependencies between nameplate attributes. The comprehensive representation vector and cross-feature tensors are fused and input into the analysis model for status classification and prediction, outputting overload risk scores, aging levels, and alarm levels, significantly enhancing intelligent recognition capabilities. The prediction results are written back to the graph, supporting status tracking and closed-loop graph updates, improving real-time performance and management efficiency.

[0068] In this embodiment, S1 specifically includes:

[0069] S11. Read the identification information on the external nameplate of each low-voltage power distribution equipment and collect six parameters: rated voltage, rated current, rated frequency, manufacturing date, installation location and insulation class.

[0070] S12. Use information acquisition tools to parse the nameplate parameters, convert the rated voltage, rated current and rated frequency values ​​into a standard format in volts, amperes and hertz, and organize the field names, numerical units and representation methods into a unified template structure and record it as a static structured field.

[0071] S13. Calculate the service life of the current equipment based on the manufacturing date field, and add the service life value as the initial aging parameter to the static structured field;

[0072] S14. Encode the installation location field to convert the descriptive geographic information into the substation, line zone or area number where the low-voltage power distribution equipment is located, and ensure that the spatial attributes can be mapped to the power distribution network topology.

[0073] S15. Connect to the operation monitoring terminal of low-voltage power distribution equipment to collect real-time operation data on current, voltage, power factor and temperature changes during operation;

[0074] S16. Add a corresponding timestamp to each piece of running data, and arrange the collected running data in chronological order to form a multi-dimensional dynamic running sequence.

[0075] This implementation method achieves standardized modeling of static attributes and dynamic behaviors by uniformly collecting nameplate parameters and operating data from low-voltage power distribution equipment. Through field parsing and unit conversion, it ensures consistency and comparability of key parameters such as rated voltage, current, and frequency. It introduces an aging dimension by estimating service life through manufacturing dates, enhancing lifespan assessment capabilities. Installation location coding structures geographical description information, strengthening the spatial mapping capability of the power distribution network. Simultaneously, it collects current, voltage, power factor, and temperature data during operation and adds timestamps to form a time-series sequence, enabling dynamic tracking of operating status. The overall process provides an accurate, unified, and time-series-complete data foundation for subsequent knowledge graph construction and analysis model input, significantly improving the reliability and scalability of equipment status analysis.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Use the unique number of each low-voltage power distribution equipment as the central node identifier, and create a corresponding equipment node to carry the static and dynamic data association relationship of the low-voltage power distribution equipment.

[0078] S22. Based on the nameplate parameters bound in the equipment nodes, create six types of nameplate parameter nodes: rated voltage, rated current, rated frequency, manufacturing date, installation location, and insulation class. Connect each parameter node to the corresponding equipment node through attribute connection edges to form a one-to-many static association structure between low-voltage power distribution equipment and nameplate parameters.

[0079] S23. Divide the collected operating data of current, voltage, power factor and temperature changes into continuous time segments with equal time intervals in chronological order. Create an operating data node for each time segment and record the dynamic status value within the time window.

[0080] S24. Connect each running data node to the corresponding device node in sequence through time relationship edges, and mark the timestamp information to form a time-series dynamic connection relationship between the device and the running status.

[0081] S25. Integrate the equipment nodes, nameplate parameter nodes, and operation data nodes into a unified multimodal knowledge graph to establish a complete equipment-nameplate-state triple relationship structure for subsequent joint modeling and feature aggregation of graph neural networks.

[0082] This implementation constructs three types of nodes—equipment, nameplate parameters, and operating status—to establish a multimodal knowledge graph structure, enabling the fusion modeling of static and dynamic information of power distribution equipment. By setting the equipment number as the central node and associating it with six types of nameplate parameter nodes, a structured static feature network is formed. Operating data is divided into equal time segments and operating data nodes are generated, establishing a time-series node chain to capture the dynamic changes of low-voltage power distribution equipment. Equipment nodes establish many-to-one associations with nameplate nodes through attribute edges and construct continuous temporal connections with operating data nodes through time relationship edges, achieving a graph representation of equipment status. This provides a unified, high-dimensional, and scalable data foundation for subsequent deep aggregation of graph neural networks and refined analysis of equipment status, significantly enhancing the expressiveness and generalization ability of status modeling.

[0083] In this embodiment, S3 specifically includes:

[0084] S31. Encode the text content and numerical fields of each device node, nameplate parameter node and operation data node in the multimodal knowledge graph. Convert the categorical fields into one-hot vectors. Subtract the preset minimum value from the rated voltage, rated current, rated frequency and manufacturing date fields respectively, and then divide by the difference between the maximum value and the minimum value to obtain normalized values ​​with a value range between 0 and 1. Finally, concatenate all categorical fields into a normalized node feature vector of the same length.

[0085] S32. In each device node, read all connected nameplate parameter nodes according to the pointing relationship of attribute edges. According to the type of nameplate parameter fields, map discrete fields to fixed-length binary vectors through one-hot encoding. Scale the values ​​of continuous fields to the [0,1] range through the max-min normalization method. Concatenate the encoding results in a preset order to generate normalized node feature vectors. Align all normalized feature vectors one by one in the dimension direction according to the element position. Add all the values ​​at the corresponding positions to obtain a summary vector. Divide the summary result by the number of nameplate parameter nodes to obtain an average vector. Generate a static structural feature representation for subsequent fusion modeling with dynamic state information.

[0086] S33. In each device node, read the connected operating data nodes in chronological order, extract four operating parameters from each operating data node: current, voltage, power factor and temperature change. Use the rated current, rated voltage, rated frequency and historical maximum temperature rise value in the nameplate parameters as normalization reference values, and perform division normalization on the current collected values ​​to obtain the four normalized values.

[0087] S34. Arrange the four normalized values ​​of all nodes in chronological order as one record per row, and arrange them row by row to form a two-dimensional time series matrix with time as the row and index as the column. This matrix is ​​used to represent the changes in the operating state at continuous time points and to generate a time series change feature matrix representing the operating state.

[0088] S35. Input the static structural feature vector and the temporal change feature matrix into the multimodal knowledge graph neural network, and perform vector dot product operation with the feature representation vector of each low-voltage power distribution equipment respectively, and calculate the matching score between the static structural features and the temporal change features.

[0089] S36. Input the matching scores of static structural features and temporal change features into the Softmax function for normalization to obtain the attention weight coefficients of static structural features and temporal change features. Perform weighted summation on the two types of features according to the corresponding attention weight coefficients to generate the fused comprehensive representation vector.

[0090] This implementation normalizes and encodes the fields of device nodes, nameplate parameter nodes, and operational data nodes, uniformly converting them into standardized node feature vectors to ensure consistent expression formats across different data types. Leveraging a multimodal knowledge graph structure, static structural feature vectors and temporal variation feature matrices are constructed from the nameplate parameter nodes and operational data nodes aggregated and connected from the device nodes, comprehensively characterizing the device's nameplate capability boundaries and real-time operational status. In the fusion encoding stage, dot product and Softmax normalization are used to calculate the matching scores of the two types of features and convert them into attention weights, achieving dynamic balanced aggregation between features. The final generated comprehensive representation vector integrates nameplate characteristics and state evolution features, providing a high-dimensional, unified, deep semantic representation for subsequent device operational status identification, significantly improving the accuracy and robustness of the analysis model.

[0091] In this embodiment, S4 specifically includes:

[0092] S41. Read the rated voltage field and manufacturing date field respectively. Obtain the maximum and minimum values ​​of the rated voltage field and manufacturing date field in all low-voltage power distribution equipment. According to the linear mapping formula between the maximum and minimum values, convert the field values ​​into normalized values. Then, concatenate the two normalized values ​​into a two-dimensional feature vector.

[0093] S42. Read the rated current field and the installation location field. Convert the rated current field to a normalized value using the same normalization formula. Perform one-hot encoding on the installation location field. Pre-set a list of all possible installation location types. Assign an independent binary dimension to each installation location. Mark the dimension corresponding to the current device's installation location as 1 and the other dimensions as 0. This generates a one-hot encoded vector with a length equal to the number of location categories. Then, concatenate the two vectors in order to form a two-dimensional feature vector.

[0094] S43. Normalize the rated voltage field and rated current field respectively and splice them to form a two-dimensional vector. Normalize the manufacturing date field and splice the insulation class field after thermal encoding to form another set of two-dimensional vectors.

[0095] S44. Input the obtained two-dimensional vector into multiple subheads in the multi-head attention mechanism. In each subhead, split the two-dimensional vector into two one-dimensional vectors. Multiply the corresponding values ​​of each dimension of the two one-dimensional vectors to obtain an intermediate vector. Sum all elements of the intermediate vector and use it as the original relevance score of the field pair in the current subhead.

[0096] S45. The scores of all field pairs in the same subhead are uniformly fed into the Softmax function for normalization to obtain the attention weights of the field pairs under the subhead. The attention weights are multiplied by the value of each dimension of the original feature vector to obtain the weighted feature representation. The weighted feature vectors output by all subheads are concatenated and input into a linear transformation network for dimension unification. Finally, the cross feature representation vector of the field pairs is output. The above operation is repeated for the four sets of field pairs. The four sets of output feature vectors together form the cross feature tensor of the nameplate parameters, which is used for subsequent fusion modeling with the running state.

[0097] This implementation normalizes and encodes the nameplate parameter field pairs, transforming each pair of fields into a standardized two-dimensional feature vector, ensuring a unified numerical expression for different parameter types. In the multi-head attention mechanism, multiple sub-heads calculate the nonlinear correlation scores between each field pair, generating weights through the Softmax function to achieve fine-grained modeling of the dependencies between fields. The weighted feature vectors are then aggregated within each sub-head and uniformly mapped to form a measurable cross-feature representation of the nameplate field interactions. This effectively captures the structural dependencies between nameplate parameters, enriches the equipment's semantic information, and provides a more discernible structural prior for operational status fusion analysis.

[0098] In this embodiment, S5 specifically includes:

[0099] S51. Arrange the comprehensive representation vector and the corresponding cross feature tensor according to the last dimension. Perform concatenation operation on each row of the two vectors according to the vector dimension order. Arrange all feature values ​​in the comprehensive representation vector by column. Then concatenate all feature values ​​in the cross feature tensor in the same order. After concatenation, generate a fusion feature vector with a length equal to the sum of the dimensions of the original two sets of vectors, which serves as the unified representation input for each device.

[0100] S52. Input the fused feature vector into the input layer of the device operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Use a linear transformation function to perform feature mapping operation on the fused feature vector. Perform matrix multiplication on the input fused feature vector and the set weight matrix. Add a bias vector with the same dimension as the output dimension, which is automatically optimized together with the weight matrix during training. Output a fixed-length hidden layer feature representation to obtain the hidden feature vector.

[0101] S53. Apply the ReLU activation function to the hidden feature vector and perform a threshold transformation element by element, setting all elements less than 0 to 0 while keeping the remaining values ​​unchanged, thereby enhancing the non-linearity of feature expression. Input the activated hidden feature vector into a multi-layer fully connected structure in sequence. Perform matrix multiplication and bias summation operations once in each layer and connect it to the ReLU activation function. In each layer, continuously reconstruct the combination relationship of features. The first layer highlights the load change related pattern in the fused features, the second layer strengthens the numerical dimension related to the equipment aging trend, and the third layer enhances the response to features such as sudden signals, boundary anomalies, and time drift. Output a feature result vector containing three continuous dimensions, representing the overload risk score, the aging degree estimate, and the abnormal signal score, respectively.

[0102] S54 inputs the feature result vector into three independent output branch structures. In the first branch, the input feature vector is multiplied and summed with a 128×1 weight vector, and then a scalar bias is added to generate a single value as an overload risk score. In the second branch, two linear mapping operations are performed in sequence. The first operation maps the feature result vector to an intermediate dimension vector to extract feature information related to the aging trend. The second operation compresses the intermediate vector into a single output value as an estimate of the degree of aging.

[0103] S55. In the third branch, after multiplying and biasing the input feature vector with the multidimensional classification weight matrix, the result vector is normalized using the Softmax function. The probability value of each alarm level is output, and the category corresponding to the highest probability is selected as the alarm priority level. The three types of operation status analysis results are output.

[0104] S55. For each device whose fusion feature vector input is the fusion of multimodal graph features and multidimensional nameplate interaction features, read the overload risk score, aging degree estimate, and alarm priority level, and concatenate the three analysis results into a structured output vector. Extract the device number field from the information table into the fusion feature vector input of the fusion multimodal graph features and multidimensional nameplate interaction features, and bind it one-to-one with the output vector to generate a structured analysis record with a unique identifier. Sort the structured analysis records of all devices according to the data acquisition timestamp, write the sorted records line by line into the analysis report file, and output the operation status analysis report.

[0105] This implementation constructs a fused feature vector reflecting the operating status and nameplate structural features by splicing and fusing a comprehensive representation vector and a cross-feature tensor. This fused feature vector is then input into an equipment operating status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Linear mapping and a multi-layer nonlinear activation structure are used to extract key risk features. The branch output modules generate overload risk scores, aging degree estimates, and alarm priority levels, respectively, achieving fine-grained risk prediction and level discrimination of equipment status. Finally, the analysis results are bound to the equipment number and sorted by time to generate a structured report, effectively supporting low-voltage distribution network equipment status assessment, risk screening, and early warning dispatch.

[0106] In this embodiment, S6 specifically includes:

[0107] S61. Establish a key-value correspondence between the unique number of the device node and the overload risk score, aging degree estimate and alarm priority level in the analysis results, and generate a result update dictionary;

[0108] S62. Locate the device node that matches the device number in the multimodal knowledge graph and read the attribute fields under the node;

[0109] S63. Write the three analysis results in the result update dictionary to the corresponding device node in the form of attribute key values, and update the overload score field, aging status field and alarm level field respectively.

[0110] S64. Mark the updated device node as "analyzed", read the time identifier corresponding to the current analysis batch, create an edge structure of type "analysis result attribution", point the edge from the device node to the batch node, and write the analysis timestamp in the edge attribute to form a structured association between nodes and record the analysis timestamp.

[0111] S65. Save the updated multimodal knowledge graph structure, output the updated multimodal knowledge graph structure, and realize the graph backwriting and structure synchronization of the prediction results.

[0112] This implementation method achieves structured synchronous write-back of analysis results by writing the overload risk score, aging estimate, and alarm priority level of each device into a multimodal knowledge graph. The method updates the attribute fields of device nodes through key-value mapping, marks node status, establishes association edges with nodes in the current analysis batch, and records the analysis timestamp, forming a multimodal knowledge graph connection structure with temporal semantics. Finally, the updated multimodal knowledge graph is output, completing the linkage update between device status analysis and the knowledge graph, providing real-time and traceable data support for intelligent operation and maintenance of the power distribution network.

[0113] Example 1:

[0114] To verify the feasibility of this invention in a real-world low-voltage power distribution network data analysis scenario, it was deployed in a power distribution operation management platform within a large industrial park in a city of a province. The platform covers 94 low-voltage power distribution devices distributed across 10 independent buildings and 4 main power supply circuits, primarily responsible for the power supply to office buildings, laboratories, and manufacturing workshops. The platform is equipped with equipment operation data acquisition terminals, capable of acquiring four core operating indicators in real time: voltage, current, power factor, and temperature, and has complete nameplate information for input into the low-voltage power distribution system.

[0115] Originally, the park mainly relied on manual methods to manage equipment nameplate information by region, and daily operational data was identified through manual table lookups, with static threshold alarms only set for some indicators. This method had several serious problems: First, the equipment nameplate information was separated from the operational data, making it difficult to form a comprehensive analysis capability; second, the alarm mechanism could only capture extreme anomalies and could not detect problems such as sudden load changes, equipment aging, and abnormal time-series trends in advance; third, there was a lack of a unified structured analysis system, and equipment status reports needed to be compiled manually, which was time-consuming and prone to errors.

[0116] After deploying this invention on the platform, a multimodal knowledge graph is constructed based on the nameplate information and runtime sequence data of each device, deeply fusing the static parameters of the nameplate with the dynamic operating status. A multimodal knowledge graph neural network is used to perform feature aggregation and cross-field modeling on each device, generating a fused expression vector, which is then input into the fusion model for operational status analysis, predicting overload risk scores, aging degree estimates, and alarm priority levels. During operation, the analysis results are automatically written back to the graph nodes, forming a structured device status archive, and analysis reports are output in batches for maintenance personnel to review. The following is a core comparison data table between the method of this invention and the platform's original manual rule system during a 60-day trial operation, in terms of status prediction capability, alarm lead time, aging trend identification, and reporting efficiency:

[0117] Table 1. Performance Comparison of the Invention Method and Traditional Method in Low-Voltage Distribution Network Operation Status Analysis Task

[0118]

[0119]

[0120] As can be seen from the comparative data in Table 1, this invention exhibits the following significant advantages in actual operation: In terms of predictive capability, the accuracy of operational status prediction has significantly improved from 72.6% to 95.1%, effectively identifying previously difficult-to-capture minor load fluctuations and early, latent anomaly signals, especially demonstrating strong robustness in nighttime light load fluctuation scenarios. Regarding response timeliness, the alarm response time has been shortened from an average of 42.7 hours to 16.3 hours, enabling the prediction of potential risks when multiple indicators show slight changes, issuing early warning information, preventing the spread of faults, and greatly improving the maintenance decision-making window.

[0121] Regarding false alarm control, traditional systems suffer from a high false alarm rate of 18.2% due to their single threshold setting. This invention, through feature fusion modeling of nameplate and operating status, effectively reduces the false alarm rate to 5.1%, a reduction of nearly 72%, significantly alleviating the burden on maintenance personnel. In terms of low-voltage power distribution equipment aging identification, this invention integrates the nameplate manufacturing date with dynamic temperature rise trends to construct an aging estimation model, achieving an accuracy rate of 93.8%, replacing the original manual inspection and estimation method and saving significant manpower. Regarding reporting efficiency and structuring, traditional systems rely on manually compiling Word or Excel documents, resulting in low efficiency and untraceable data. This invention automatically outputs structured report files daily, with unified field binding and direct data import for analysis, greatly improving data asset management capabilities.

[0122] In summary, this invention not only effectively solves key problems in low-voltage distribution network operation, such as delayed status identification, difficulty in early warning of aging low-voltage distribution equipment, high false alarm rates, and low reporting efficiency, but also brings significant improvements in data structuring, predictive intelligence, and operation and maintenance automation, possessing the potential for widespread application. Further expansion to substations, transformer boxes, and city-level distribution management platforms will further enhance the intelligent perception and proactive safety capabilities of the power system.

[0123] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for analyzing low-voltage distribution network data systems by fusing model and nameplate information, characterized in that, Includes the following steps: S1. Collect the nameplate parameters of each low-voltage power distribution equipment, organize the nameplate parameters into structured static fields, and record the operation data during operation in chronological order as a multi-dimensional dynamic operation sequence. S2. Use the structured static fields of each low-voltage power distribution equipment and nameplate parameters, as well as the multidimensional dynamic runtime sequence of the operating data, as different types of nodes in the multimodal knowledge graph to construct a multimodal knowledge graph; S3. Use a multimodal knowledge graph neural network to jointly learn the node information in the multimodal knowledge graph, perform feature aggregation processing on the device nodes through the multimodal knowledge graph neural network, and deeply fuse the static structural features in the nameplate parameters with the temporal change features in the operating data to generate a comprehensive representation vector; S4. Extract four sets of static field pairs from the nameplate parameters as cross-feature groups. Through the cross-feature modeling structure of the multi-head attention mechanism, calculate the nonlinear relationship between the four sets of field pairs, extract the structural dependency features between the nameplate parameters, and output the cross-feature tensor of the nameplate parameters. S5. Concatenate the comprehensive representation vector and the cross-feature tensor, and input them into the equipment operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Classify or regress the operation status, and output the overload risk score, aging degree estimate and alarm priority level. S6. Synchronously write the prediction results back to the corresponding device nodes in the multimodal knowledge graph and update the multimodal knowledge graph structure.

2. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S1 specifically includes: S11. Read the identification information on the external nameplate of each low-voltage power distribution equipment and collect six parameters: rated voltage, rated current, rated frequency, manufacturing date, installation location and insulation class. S12. Use information acquisition tools to parse the nameplate parameters, convert the rated voltage, rated current and rated frequency values ​​into a standard format in volts, amperes and hertz, and organize the field names, numerical units and representation methods into a unified template structure and record it as a static structured field. S13. Calculate the service life of the current equipment based on the manufacturing date field, and add the service life value as the initial aging parameter to the static structured field; S14. Encode the installation location field to convert the descriptive geographic information into the substation, line zone, or area number where the low-voltage power distribution equipment is located. S15. Connect to the operation monitoring terminal of low-voltage power distribution equipment to collect real-time operation data on current, voltage, power factor and temperature changes during operation; S16. Add a corresponding timestamp to each piece of running data, and arrange the collected running data in chronological order to form a multi-dimensional dynamic running sequence.

3. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S2 specifically includes: S21. Use the unique number of each low-voltage power distribution device as the central node identifier to create the corresponding device node; S22. Based on the nameplate parameters bound in the equipment nodes, create six types of nameplate parameter nodes: rated voltage, rated current, rated frequency, manufacturing date, installation location, and insulation class. Connect each parameter node to the corresponding equipment node through attribute connection edges to form a one-to-many static association structure between low-voltage power distribution equipment and nameplate parameters. S23. Divide the collected operating data of current, voltage, power factor and temperature changes into continuous time segments with equal time intervals in chronological order. Create an operating data node for each time segment and record the dynamic status value within the time window. S24. Connect each running data node to the corresponding device node in sequence through time relationship edges, and mark the timestamp information to form a time-series dynamic connection relationship between the device and the running status. S25. Integrate equipment nodes, nameplate parameter nodes, and operation data nodes into a unified multimodal knowledge graph.

4. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S3 specifically includes: S31. Encode the text content and numerical fields of each device node, nameplate parameter node and operation data node in the multimodal knowledge graph. Convert the categorical fields into one-hot vectors. Subtract the preset minimum value from the rated voltage, rated current, rated frequency and manufacturing date fields respectively, and then divide by the difference between the maximum value and the minimum value to obtain normalized values ​​with a value range between 0 and 1. Finally, concatenate all categorical fields into a normalized node feature vector of the same length. S32. In each device node, read all connected nameplate parameter nodes according to the pointing relationship of attribute edges. According to the type of nameplate parameter fields, map discrete fields to fixed-length binary vectors through one-hot encoding. Scale the values ​​of continuous fields to the [0,1] range through the max-min normalization method. Concatenate the encoding results in a preset order to generate normalized node feature vectors. Align all normalized feature vectors one by one in the dimension direction according to the element position. Add all the values ​​at the corresponding positions to obtain the summary vector. Divide the summary result by the number of nameplate parameter nodes to obtain the average vector and generate a static structural feature representation. S33. In each device node, read the connected operating data nodes in chronological order, extract four operating parameters from each operating data node: current, voltage, power factor and temperature change. Use the rated current, rated voltage, rated frequency and historical maximum temperature rise value in the nameplate parameters as normalization reference values, and perform division normalization on the current collected values ​​to obtain the four normalized values. S34. Arrange the four normalized values ​​of all nodes in chronological order as one record per row, and arrange them row by row to form a two-dimensional time series matrix with time as the row and index as the column, to generate a time series change feature matrix representing the running status. S35. Input the static structural feature vector and the temporal change feature matrix into the multimodal knowledge graph neural network, and perform vector dot product operation with the feature representation vector of each low-voltage power distribution equipment respectively, and calculate the matching score between the static structural features and the temporal change features. S36. Input the matching scores of static structural features and temporal change features into the Softmax function for normalization to obtain the attention weight coefficients of static structural features and temporal change features. Perform weighted summation on the two types of features according to the corresponding attention weight coefficients to generate the fused comprehensive representation vector.

5. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S4 specifically includes: S41. Read the rated voltage field and manufacturing date field respectively. Obtain the maximum and minimum values ​​of the rated voltage field and manufacturing date field in all low-voltage power distribution equipment. According to the linear mapping formula between the maximum and minimum values, convert the field values ​​into normalized values. Then, concatenate the two normalized values ​​into a two-dimensional feature vector. S42. Read the rated current field and the installation location field. Use the same normalization formula to convert the rated current field into a normalized value. Perform one-hot encoding on the installation location field to generate a one-hot encoded vector with a length equal to the number of location categories. Then concatenate the two vectors in order to form a two-dimensional feature vector. S43. Normalize the rated voltage field and rated current field respectively and splice them to form a two-dimensional vector. Normalize the manufacturing date field and splice the insulation class field after thermal encoding to form another set of two-dimensional vectors. S44. Input the obtained two-dimensional vector into multiple subheads in the multi-head attention mechanism. In each subhead, split the two-dimensional vector into two one-dimensional vectors. Multiply the corresponding values ​​of each dimension of the two one-dimensional vectors to obtain an intermediate vector. Sum all elements of the intermediate vector and use it as the original relevance score of the field pair in the current subhead. S45. The scores of all field pairs in the same subhead are uniformly fed into the Softmax function for normalization to obtain the attention weights of the field pairs under the subhead. The attention weights are multiplied by each dimension of the original feature vector to obtain the weighted feature representation. The weighted feature vectors output by all subheads are concatenated and fed into a linear transformation network for dimension unification. Finally, the cross feature representation vectors of the field pairs are output. The above operation is repeated for the four sets of field pairs in turn. The four sets of output feature vectors together form the cross feature tensor of the nameplate parameters.

6. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S5 specifically includes: S51. Arrange the comprehensive representation vector and the corresponding cross feature tensor according to their last dimension. Perform concatenation operation on each row of the two vectors according to the order of their vector dimensions. Arrange all the feature values ​​in the comprehensive representation vector by column. Then, concatenate all the feature values ​​in the cross feature tensor in the same order. After concatenation, generate a fused feature vector with a length equal to the sum of the dimensions of the original two sets of vectors. S52. Input the fused feature vector into the input layer of the device operation status analysis model that integrates multimodal graph features and multidimensional nameplate interaction features. Use a linear transformation function to perform feature mapping operation on the fused feature vector. Perform matrix multiplication on the input fused feature vector and the set weight matrix. Add a bias vector with the same dimension as the output dimension, which is automatically optimized together with the weight matrix during training. Output a fixed-length hidden layer feature representation to obtain the hidden feature vector. S53. Apply the ReLU activation function to the hidden feature vector and perform threshold transformation element by element. Then, input the activated hidden feature vector into a multi-layer fully connected structure. Perform matrix multiplication and bias summation operations once in each layer and connect it to the ReLU activation function. Continuously reconstruct the combination relationship of features in each layer and output the feature result vector. S54 inputs the feature result vector into three independent output branch structures. In the first branch, the input feature vector is multiplied and summed with a 128×1 weight vector, and then a scalar bias is added to generate a single value as an overload risk score. In the second branch, two linear mapping operations are performed in sequence. The first operation maps the feature result vector to an intermediate dimension vector to extract feature information related to the aging trend. The second operation compresses the intermediate vector into a single output value as an estimate of the degree of aging. S55. In the third branch, after multiplying and biasing the input feature vector with the multidimensional classification weight matrix, the result vector is normalized using the Softmax function. The probability value of each alarm level is output, and the category corresponding to the highest probability is selected as the alarm priority level. The three types of operation status analysis results are output. S55. For each device whose fusion feature vector input is the fusion of multimodal graph features and multidimensional nameplate interaction features, read the overload risk score, aging degree estimate, and alarm priority level, and concatenate the three analysis results into a structured output vector. Extract the device number field from the information table into the fusion feature vector input of the fusion multimodal graph features and multidimensional nameplate interaction features, and bind it one-to-one with the output vector to generate a structured analysis record. Sort the structured analysis records of all devices according to the data acquisition timestamp, write the sorted records line by line into the analysis report file, and output the operation status analysis report.

7. The low-voltage distribution network data system analysis method based on the fusion of model and nameplate information according to claim 1, characterized in that, S6 specifically includes: S61. Establish a key-value correspondence between the unique number of the device node and the overload risk score, aging degree estimate and alarm priority level in the analysis results, and generate a result update dictionary; S62. Locate the device node that matches the device number in the multimodal knowledge graph and read the attribute fields under the node; S63. Write the three analysis results in the result update dictionary to the corresponding device node in the form of attribute key values, and update the overload score field, aging status field and alarm level field respectively. S64. Mark the updated device node as "analyzed", read the timestamp corresponding to the current analysis batch, create an edge structure of type "analysis result attribution", point the edge from the device node to the batch node, and write the analysis timestamp into the edge attribute to record the analysis timestamp. S65. Save the updated multimodal knowledge graph structure and output the updated multimodal knowledge graph structure.

Images