Autonomous learning construction method and system for engineering SCD check model

Abstract

The invention discloses an autonomous learning construction method and system of an engineering SCD check model. A graph structure is extracted from positive and negative samples of a historical SCD configuration file to be embedded and expressed, and text features are extracted from historical unstructured text data. A generator is designed to process positive sample data and text data, two parallel processing networks are utilized to capture graph structure features and description features respectively, and SCD is fused and output to generate samples. A multi-task discrimination module is designed in a discriminator architecture, positive and negative samples are learned, and equipment topology accuracy, text description fitness and configuration errors are analyzed for SCD generation samples. Iteratively training and optimizing a generator and a discriminator according to an analysis result, and calling the newly output discriminator for checking when a new SCD configuration file is input; and bringing the new SCD configuration file and the check result into collection of sample data, and executing incremental learning. According to the invention, the accuracy, efficiency and self-learning ability of SCD checking are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent substation control technology, specifically relating to an autonomous learning construction method and system for an engineering SCD verification model. Background Technology

[0002] System Configuration Description (SCD) verification in engineering projects typically refers to the process of checking and verifying system configuration files or data in power systems or other engineering projects. Its purpose is to ensure that the system meets predetermined technical requirements during the design and implementation phases, operates normally, and is consistent with relevant standards and specifications at different stages of the project. Verification content includes data consistency, compatibility, logical consistency, security, and performance. Through these verifications, engineers can promptly identify and correct potential design flaws or configuration problems, reducing risks during implementation and subsequent operation and maintenance costs.

[0003] Existing technologies primarily employ standardization and automated comparison methods to improve the standardization and quality of SCD files for smart substations. For example, patent application CN119960847A converts the SCD file content into a computable byte stream through standardization processing and dictionary mapping, and then performs efficient and accurate difference comparison using CRC32 checksums. Another example is patent application CN118210758A, which uses a keyword combination algorithm to match and standardize the actual virtual terminals in the engineering SCD with standard virtual terminals. The standardized virtual circuit configurations are then compiled into a library of circuits awaiting activation. Pre-activation circuits are marked by statistical frequency, and then compared with the library of circuits awaiting activation.

[0004] While the aforementioned existing technologies have improved the reliability and operation and maintenance efficiency of the secondary system of smart substations to some extent, they still have shortcomings: the learning ability of existing verification models is limited by the defined syntax rules and keyword system, making it difficult to understand the diversity and context of engineering semantics, resulting in insufficient generalization ability when facing new equipment and non-standard designs; at the same time, confirming whether it is standard by statistical frequency lacks a verification closed loop related to the correctness of power grid business functions. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a self-learning construction method and system for an engineering SCD verification model. Graph structure embeddings are extracted from positive and negative samples of historical SCD configuration files, and text features are extracted from historical unstructured text data. A generator is designed to process positive sample data and text data, utilizing two parallel processing networks to capture graph structure features and descriptive features respectively, and then fusing them to output SCD generated samples. A multi-task discrimination module is designed in the discriminator architecture to learn from positive and negative samples and analyze the SCD generated samples for device topology accuracy, text description fit, and configuration errors. The generator and discriminator are iteratively trained and optimized based on the analysis results. When a new SCD configuration file is input, the latest output discriminator is called for verification; the new SCD configuration file and verification results are incorporated into the collection of sample data for incremental learning.

[0006] The first aspect of this application discloses an autonomous learning construction method for an engineering SCD verification model, which adopts the following technical solution: Sample data collection and preprocessing; extracting structured information from historical SCD configuration files and converting it into graph structure embedding representation; collecting historical unstructured text data and performing text embedding representation; the historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively; Based on power grid demand and standard design, a generator is used to process positive sample data and text data. Two parallel processing networks are designed in the generator architecture to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation, respectively. The graph structure features and description features are fused to output SCD generated samples. This is used to design a discriminator based on the SCD structure; a multi-task discriminator module is designed in the discriminator architecture to learn positive and negative samples and perform analysis on the SCD-generated samples; including analyzing the accuracy of device topology in the SCD-generated samples, judging the fit of text descriptions, and identifying configuration errors; Iterative adversarial training and dynamic updates; the generator and discriminator are optimized through iterative training, and when a new SCD configuration file is input, the latest output discriminator is called for verification; the new SCD configuration file and verification results are incorporated into the collection of sample data to perform incremental learning.

[0007] Furthermore, the generator is designed including: The parallel processing network is divided into a graph data processing network and a text data processing network; The graph data processing network is used to extract internal graph structure features of the power grid from positive samples. When using a graph convolutional network to aggregate node features, an adaptive adjacency matrix mechanism is introduced to update the adjacency matrix of each training step of the graph convolutional network, and this is controlled by a regularization mechanism.

[0008] Furthermore, in the multi-task discrimination module, analyzing the accuracy of device topology includes: Extract the graph structure from the SCD generated file as the generated topology, and call the graph structure in the positive sample as the standard topology; Calculate the number of topology elements that are exactly the same as the standard topology. and the total number of topological elements in the standard topology. ;by and The ratio is used to calculate the topology accuracy score.

[0009] Furthermore, in the multi-task discrimination module, the discrimination of text description fit includes: The text description field of the device is extracted from the SCD generated file, and the standard description is extracted from the positive sample. For a single device, the matching degree of each field between the text description field and the standard description is calculated, and then weighted to obtain the text description fit score of the single device. The average of the text description fit scores for all devices in the SCD generated file is used to obtain the text description fit score for the SCD generated file.

[0010] Furthermore, the multi-task discrimination module identifies configuration errors, including: Learn correct and incorrect configurations based on positive and negative samples, and train a classifier; The classifier is used to analyze the SCD generated file to generate a configuration error score; the configuration error score is represented as the weighted sum of the mean probability of error of all device parameters, the mean probability of error of topology, and the mean probability of error of device function in the SCD generated file output by the classifier.

[0011] Furthermore, the generator and discriminator are optimized through iterative training, including: The generator adjusts its loss function based on the topology accuracy score, text description fit score, and configuration error score output by the discriminator. The adjustment method is to add a quality score term based on the discriminator output to the generator's loss function on top of the adversarial loss term. The quality scoring items are the topological accuracy of the SCD-generated text, the text description fit, and the weighted sum of configuration errors.

[0012] Furthermore, in the calculation of the quality score items, topology accuracy is taken as 1 minus the topology accuracy score; text description fit is taken as 1 minus the text description fit score.

[0013] The second aspect of this application discloses an autonomous learning construction system for an engineering SCD verification model, which implements the autonomous learning construction method described in the first aspect of this application. The system includes: The sample data acquisition module is used for sample data collection and preprocessing; it extracts structured information from historical SCD configuration files and converts it into graph structure embedding representation; it collects historical unstructured text data and performs text embedding representation; the historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively. A generator design module is used to design a generator based on power grid requirements and standards to process positive sample data and text data. Two parallel processing networks are designed in the generator architecture to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation, respectively. The graph structure features and description features are fused to output SCD generated samples. Discriminator design module; used to design a discriminator based on the SCD structure; in the discriminator architecture, a multi-task discrimination module is designed to learn positive and negative samples and perform analysis on SCD-generated samples; including analyzing the accuracy of device topology in SCD-generated samples, judging the fit of text descriptions, and identifying configuration errors; The adversarial optimization module is used for iterative adversarial training and dynamic updates. It optimizes the generator and discriminator through iterative training; when a new SCD configuration file is input, the latest output discriminator is called for verification; the new SCD configuration file and verification results are incorporated into the collection of sample data for incremental learning. The beneficial effects of this invention are that, compared with the prior art, 1. This application improves the accuracy and reliability of SCD configuration files. By using graph convolutional networks and text embedding representations, combined with power grid topology features and equipment description features, the generated SCD configuration files more accurately reflect power grid design requirements. The multi-task discrimination module effectively identifies and verifies potential problems in the configuration by analyzing the accuracy of equipment topology, the fit of text descriptions, and configuration errors in the generated SCD file. This automated verification and feedback mechanism helps reduce manual intervention, improves the accuracy and reliability of configuration files, helps reduce power grid operation problems caused by configuration errors, and ensures the correctness and consistency of power grid equipment and topology.

[0014] 2. This application achieves adaptive optimization of the model through incremental learning. It employs iterative adversarial training and dynamic updates, adjusting the generator's loss function based on the discriminator's output to ensure continuous learning and optimization by both the generator and discriminator. Whenever a new SCD profile is input, the latest discriminator verifies it and feeds the verification result back to the sample dataset. This enables the model to adapt to changes in grid demand, such as equipment additions and load variations, ensuring that the generated SCD profile dynamically adjusts and adapts to the new grid environment. Simultaneously, through incremental learning, the model can self-optimize after each update, improving the accuracy and robustness of subsequent generation and verification. Attached Figure Description

[0015] Figure 1 The following is an execution flowchart of the autonomous learning construction method described in the embodiment. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of this invention.

[0017] As an embodiment of this application, a specific implementation method for an autonomous learning construction method of an engineering SCD verification model is disclosed. The execution flow of the method embodiment is as follows: Figure 1 .

[0018] S1: As one implementation method, sample data collection and preprocessing are performed. Structured information is extracted from historical SCD configuration files and transformed into graph structure embedding representation. Historical unstructured text data is collected and text embedding representation is performed. The historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively.

[0019] 1.1: In a further implementation, historical SCD configuration file data needs to be collected. The data source is not limited, but primarily focuses on the systems used in the deployment scenario. Examples include power system design departments, smart substation operation and management systems, and historical power grid project documents.

[0020] The collected SCD configuration file data includes two categories: standard configuration files and non-standard configuration files.

[0021] Standard configuration files refer to SCD configuration files that meet the requirements of existing power grid design standards, equipment models, design specifications, etc. They are usually design blueprints or configuration lists that conform to the specifications. Non-standard configuration files refer to SCD configuration files that do not fully meet the standard design requirements. These files may contain errors, omissions, or configuration items that are incompatible with the current standards.

[0022] During data collection, existing engineering documents, design libraries, equipment management systems, etc., can be crawled through automated scripts or interfaces, and historical SCD configuration files can be labeled: standard configuration files are labeled as positive samples; non-standard configuration files are labeled as negative samples.

[0023] In one optional implementation, data cleaning and format standardization are performed to ensure data quality. This includes deleting invalid and duplicate data, handling null and missing values, and ensuring data integrity. It may also include standardizing the format of SCD files from different sources, handling inconsistencies such as different units and date formats in the data, and ensuring data consistency and comparability.

[0024] 1.2: In a further implementation, structured information is extracted from historical SCD configuration files and transformed into a graph structure embedded representation.

[0025] SCD configuration files typically consist of multi-level structured data, including the identifiers of power grid equipment, connection relationships, and equipment parameters. These files are parsed to extract their structured information. Commonly extracted information includes: the ID, type, location, and function of the power grid equipment; the connection relationships between equipment, such as current, power, and signal transmission; and configuration parameters, equipment status, and protocols.

[0026] In one optional implementation, the SCD file is parsed layer by layer using an XML parser, a JSON parser, or custom parsing rules to obtain structured information about the devices and their connections.

[0027] The extracted structured information is transformed into a graph structure representation: each device is regarded as a node, and the node attributes are device information (such as device ID, device type, device status, etc.); the connection relationship between devices is regarded as an edge, wherein the connection relationship includes but is not limited to current flow, signal transmission, control cables, etc.

[0028] In one specific implementation, an embedded representation of the graph structure is generated using algorithms such as graph convolutional networks.

[0029] 1.3: In a further implementation, historical unstructured text data is collected and text embedding representation is performed.

[0030] In addition to historical SCD configuration files, it is also necessary to collect historical unstructured text data related to power grid design. This includes, but is not limited to: project design documents, technical specifications, equipment manuals, operation guides, power grid fault analysis reports, and operation and maintenance records. These text data can originate from power system document libraries, project management systems, and operation and maintenance logs. The collected text data also needs to undergo preprocessing operations such as cleaning, noise reduction, and deduplication to ensure text quality.

[0031] The collected text data undergoes word segmentation and stop removal, followed by word embedding modeling to embed the text, converting each word into a low-dimensional, dense vector representation. A sentence-based pre-trained model is then used to further embed the text, obtaining the sentence's vector representation. Through these steps, unstructured text data is transformed into a computable embedding representation.

[0032] In further optional and non-mandatory implementations, the text embedding model (including word embedding and sentence embedding) can use basic methods such as Word2Vec, GloVe, BERT, or their variants. Existing libraries can be called in PyCharm; mature existing technologies will not be elaborated upon here.

[0033] S2: As one implementation method of the embodiment, a generator is designed based on power grid demand and standards to process positive sample data and text data; the generator architecture is designed with two parallel processing networks to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation respectively; the graph structure features and description features are fused to generate an SCD configuration file, and the generator is trained by combining historical SCD configuration files as training samples, and the generated SCD samples are output.

[0034] 2.1: In a further implementation, the graph structure embedding representation extracted from historical SCD profiles (using positive samples here) is used as input to the first processing network. The first processing network employs a graph convolutional network to aggregate and pass the topology, connectivity, and device attributes from the graph structure embedding representation. Specifically: During the training of a graph neural network, the features of each node are iteratively updated using the features of its neighboring nodes, thereby capturing the structural information of the entire power grid. Each layer of the graph convolutional network updates the node representation based on the node's adjacency information, gradually capturing the connection relationships between power grid devices.

[0035] In GCN, an adjacency matrix is ​​required. This is used to represent the connection relationships between devices. Among them, Represents a node and The adjacency matrix establishes the connections between elements. Normalization is applied to the adjacency matrix to ensure data stability and convergence.

[0036] However, in the task of generating the SCD configuration file for the power grid, the connection relationships between power grid devices are not static and may change depending on factors such as device operating status, load demand, and fault recovery. Conventional GCN methods assume the stability of the connection relationships between devices throughout the process by using a fixed adjacency matrix. This may be effective for some static power grid scenarios, but it is likely to be a limitation for power grid designs with dynamic demands and changes.

[0037] Therefore, in a further specific implementation, this application introduces an adaptive adjacency matrix, enabling the adjacency matrix of each layer of the GCN to be dynamically updated, thereby adjusting the connection relationships between nodes according to data. Specifically: Set the initial adjacency matrix During each round of training, the adjacency matrix is ​​processed by a trainable mapping function. And this is controlled through a regularization mechanism. It is represented as: ; In the formula, and These are the adjacency matrices for the current training step and the next training step, respectively. The mapping function called for the adjacency matrix of the current training step; This is the regularization coefficient, used to control the strength of the regularization term; This is a regularization term for the adjacency matrix, ensuring that the adjacency matrix does not deviate excessively from the standard structure during adjustment. In a further specific implementation, L2 regularization can be used: ; This represents the regularization term of the adjacency matrix at the current training step. It is an L2 norm.

[0038] Furthermore, the mapping function The form is ; For normalization function, It is a non-linear activation function; The node feature matrix is ​​defined in this application. Each row represents a feature vector of a power grid device, including voltage, current, power, device type, etc. This is the weight matrix.

[0039] After processing by the improved graph convolutional network, the output includes the internal graph structure features of the power grid, including node embedding vectors and matrices, which are used for the generation of subsequent configuration files.

[0040] 2.2: In a further embodiment, the text embedding representation extracted from the unstructured text data is used as input to the generator's second processing network. The second processing network employs a deep neural network or a convolutional neural network, combined with a self-attention mechanism, to output descriptive features of the power grid equipment.

[0041] 2.3: In a further implementation, the internal graph structure features of the power grid and the description features of the power grid equipment of multiple sets of positive and negative samples output by the two parallel processing networks are fused to generate an SCD configuration file.

[0042] The internal graph structure features of the power grid and the descriptive features of power grid equipment are fused and processed into a feature vector. And through a fully connected network or sequence generation model, an SCD configuration file containing all the necessary fields is generated.

[0043] In one alternative implementation, the strategy for feature fusion may include splicing, weighting, or attention mechanisms.

[0044] S3: As one implementation method of the embodiment, a discriminator is designed to learn positive and negative samples simultaneously; in the discriminator architecture, a multi-task discriminator module is designed to perform parallel analysis on the generated SCD samples; including analyzing the accuracy of device topology in the generated SCD samples, judging the fit of text descriptions, and identifying configuration errors.

[0045] In a further embodiment, the multi-task discrimination module is used to analyze and discriminate the topological accuracy, text description fit, and configuration errors of the SCD-generated samples.

[0046] The discriminator takes the SCD generated file from step 2 as input and outputs a score indicating whether the generated file conforms to standards, including topology accuracy score, text description fit score, and configuration error score. Based on the score, the generator receives feedback to further optimize the generated configuration file. Specifically: 3.1: Topological accuracy; The graph structure extracted from the SCD generated file is called the generated topology, which includes device nodes, connection relationships, and adjacency matrix. The graph structure in the standard configuration file (i.e., the positive sample) is called the standard topology. The topological elements in a graph structure are defined as power grid devices and their connections, namely nodes (each device in the power grid, such as transformers, switches, generators, etc.) and edges (connections between power grid devices, such as current, voltage, power transmission, etc.).

[0047] Determine if each topology element in the generated topology is consistent with the standard topology. Check if each device in the generated topology matches a device in the standard topology, and check if the connections between devices in the generated topology conform to the connection relationships in the standard topology.

[0048] Calculate the number of topology elements that are exactly the same as the standard topology. and the total number of topological elements in the standard topology. Then the topological accuracy score The calculation method is as follows: ; The meaning of topology accuracy score is as follows: when the generated topology perfectly matches the standard topology, the score is 1, indicating that there is no topology error. If the topology is inconsistent, the score will be lower than 1, indicating that there is an error or inconsistency in the generated topology. The lower the score value, the more serious the error.

[0049] In one optional implementation, node consistency and edge consistency checks are performed by directly comparing the adjacency matrix. For example, if nodes in the generated topology... and nodes If the connection relationship is consistent with the connection relationship in the standard topology, then the edge (connection) is considered correct.

[0050] 3.2: Text description fit; Each device in the SCD configuration file has its corresponding structured fields, such as device signal, rated voltage, rated power, and functional description.

[0051] Extract the text description field of the device from the SCD generated file. For example, the text description of the device may be: "Transformer model X123, rated voltage 10kV, capacity 200MVA"; Extract standard descriptions from standard SCD configuration files, such as "Transformer model X123, rated voltage 10kV, rated power 200MVA, used for step-up".

[0052] For each item in the structured field, it is first converted into word vectors or embedding vectors using text embedding methods. Then, the similarity between each item and the generated description and the standard description is calculated. As an example, the transformer description field is used below.

[0053] The similarity between the text vectors of "Model X123" in the generated description and the "Model [Device Model]" part in the standard description is calculated as the device model matching degree. ; The similarity between the text vectors of "rated voltage 10kV" in the generated description and the "rated voltage [rated voltage]" part in the standard description is calculated as the rated voltage matching degree. ; The similarity between the text vectors of "rated power 200MVA" in the generated description and the "rated power [rated power]" part in the standard description is calculated as the rated power matching score. ; Calculate the similarity between the text vectors of "for boost" in the generated description and "for [functional description]" in the standard description, and use this as the functional description matching score. ; Weight the matching degree of all description fields to obtain the device. Text description fit score This is represented as: ; In the formula, The weights corresponding to the matching degree of each field.

[0054] Extending this to other devices, it can be expressed as ; For field indexing, The number of fields. Then the text description fit score for the entire SCD file. Represented as ; For device indexing, The number of devices generated in the SCD file.

[0055] 3.3 Configuration error scoring; To effectively learn during the configuration error scoring process, features must first be extracted from positive and negative samples, and a classifier must be trained to enable the discrimination module to identify configuration errors. Configuration errors in negative samples include, but are not limited to: parameters that do not conform to power grid design standards, such as mismatched equipment power and rated voltage; topology errors, such as unreasonable connections, closed loops, or disconnections; and inconsistencies between equipment function and configuration, such as equipment type and functional description.

[0056] In one optional implementation, correct configurations from positive samples are also learned. Positive samples (standard configuration files) represent configurations that conform to specifications. The discriminator needs to learn the correct configuration features from these positive samples to compare with the incorrect features from negative samples. Correct configuration features corresponding to configuration errors include, but are not limited to, parameter consistency, topology consistency, and device functionality consistency. The classifier is trained based on the configuration features from positive and negative samples, and supervised learning can be invoked.

[0057] The trained classifier is used to analyze the SCD-generated files and generate configuration error scores. ; In the formula, Scoring for configuration errors; , and These represent the mean error probabilities of all device parameters, the mean error probabilities of topology, and the mean error probabilities of device functions in the SCD generation file output by the classifier, respectively. , and These are the weighting coefficients for the mean values ​​of each item.

[0058] S4: As one implementation method of the embodiment, iterative adversarial training and dynamic updates are performed; the generator and discriminator are optimized through iterative training; when a new SCD configuration file is input, the latest output discriminator is called for verification; the new SCD configuration file and verification results are incorporated into the collection of sample data, and incremental learning is performed.

[0059] 4.1: In a further implementation, the generator aims to "deceive" the discriminator, preventing it from correctly distinguishing between the generated SCD configuration file and the actual standard configuration file. The discriminator's goal is to identify whether the generated file conforms to the standard configuration as accurately as possible. Therefore, the score output by the discriminator directly affects the generator's optimization process. Specifically: The discriminator outputs topology accuracy score, text description fit score, and configuration error score through the multi-task discrimination module. The generator's loss function is then adjusted as follows: ; In the above formula, The loss function for the generator; To counteract the loss, Indicates the SCD generated file; The discriminator considers the generated sample It is the probability of being a real sample; , and These represent the weights of the impact of device topology accuracy, text description fit, and configuration errors on generator optimization, respectively.

[0060] The generator and discriminator are trained in an adversarial manner. The generator acts as the "discriminator" by attempting to generate SCD profiles that better conform to the standards; the discriminator, in turn, "challenges" the generator by correctly classifying the generator's output and the true standard profiles. During training, the generator and discriminator alternately update, continuously improving each other's performance.

[0061] 4.2: In a further implementation, a new SCD profile will be generated when the grid demand changes (such as grid expansion, new equipment access, load changes, etc.).

[0062] The new SCD configuration file is input into the latest optimized discriminator for verification. The discriminator analyzes the configuration file based on scoring criteria such as topology accuracy, text description fit, and configuration errors, and outputs the corresponding score.

[0063] 4.3: In a further implementation, the verification results are used together with the newly input SCD configuration file as new training samples.

[0064] If the new SCD profile conforms to the specifications and all scores are in the high score range (e.g., [0.85, 1]), it is considered a positive sample. If any score is below 0.85, the new SCD profile is considered to have a problem (e.g., topology error or misconfiguration) and is considered a negative sample.

[0065] These newly input SCD configuration files and their verification results are added to the positive and negative sample sets in the training dataset to expand the training samples.

[0066] In a further implementation, incremental learning is performed based on the newly added sample dataset to optimize the generator and discriminator. Specifically: The generator is trained using new sample data. The generator adjusts its generation strategy using these new positive and negative samples, making the generated SCD profile more compliant with the specifications.

[0067] The discriminator was further optimized using new verification data, enabling it to better distinguish between compliant and non-compliant configuration files.

[0068] As an optional implementation method, incremental learning uses online learning or mini-batch learning methods. Instead of retraining the entire model, it fine-tunes the existing model with new sample data, allowing it to adapt to new data while retaining previous knowledge.

[0069] As an embodiment of this application, a self-learning construction system for an engineering SCD verification model is disclosed. Employing the specific implementation method described above, the system includes: The sample data acquisition module is used for sample data collection and preprocessing; it extracts structured information from historical SCD configuration files and converts it into graph structure embedding representation; it collects historical unstructured text data and performs text embedding representation; the historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively. A generator design module is used to design a generator based on power grid requirements and standards to process positive sample data and text data. Two parallel processing networks are designed in the generator architecture to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation, respectively. The graph structure features and description features are fused to output SCD generated samples. Discriminator design module; used to design a discriminator based on the SCD structure; in the discriminator architecture, a multi-task discrimination module is designed to learn positive and negative samples and perform analysis on SCD-generated samples; including analyzing the accuracy of device topology in SCD-generated samples, judging the fit of text descriptions, and identifying configuration errors; The adversarial optimization module is used for iterative adversarial training and dynamic updates. It optimizes the generator and discriminator through iterative training. When a new SCD configuration file is input, the latest output discriminator is called for verification. The new SCD configuration file and verification results are incorporated into the collection of sample data to perform incremental learning.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A method for constructing an autonomous learning model for engineering SCD verification, characterized in that, include: Sample data collection and preprocessing; extracting structured information from historical SCD configuration files and converting it into graph structure embedding representation; collecting historical unstructured text data and performing text embedding representation; the historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively; Based on power grid demand and standard design, a generator is used to process positive sample data and text data. Two parallel processing networks are designed in the generator architecture to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation, respectively. The graph structure features and description features are fused to output SCD generated samples. This is used to design a discriminator based on the SCD structure; a multi-task discriminator module is designed in the discriminator architecture to learn positive and negative samples and perform analysis on the SCD-generated samples; This includes analyzing the accuracy of device topology in SCD-generated samples, determining the relevance of text descriptions, and identifying configuration errors; Iterative adversarial training and dynamic updates; the generator and discriminator are optimized through iterative training, and when a new SCD configuration file is input, the latest output discriminator is called for verification; the new SCD configuration file and verification results are incorporated into the collection of sample data to perform incremental learning.

2. The autonomous learning construction method for an engineering SCD verification model according to claim 1, characterized in that, Design the generator, including: The parallel processing network is divided into a graph data processing network and a text data processing network; The graph data processing network is used to extract internal graph structure features of the power grid from positive samples. When using a graph convolutional network to aggregate node features, an adaptive adjacency matrix mechanism is introduced to update the adjacency matrix of each training step of the graph convolutional network, and this is controlled by a regularization mechanism.

3. The autonomous learning construction method for an engineering SCD verification model according to claim 1, characterized in that, The multi-task discrimination module analyzes the accuracy of device topology, including: Extract the graph structure from the SCD generated file as the generated topology, and call the graph structure in the positive sample as the standard topology; Calculate the number of topology elements that are exactly the same as the standard topology. and the total number of topological elements in the standard topology. ;by and The ratio is used to calculate the topology accuracy score.

4. The autonomous learning construction method for an engineering SCD verification model according to claim 1, characterized in that, The multi-task discrimination module determines the relevance of the text description, including: The text description field of the device is extracted from the SCD generated file, and the standard description is extracted from the positive sample. For a single device, the matching degree of each field between the text description field and the standard description is calculated, and then weighted to obtain the text description fit score of the single device. The average of the text description fit scores for all devices in the SCD generated file is used to obtain the text description fit score for the SCD generated file.

5. The autonomous learning construction method for an engineering SCD verification model according to claim 1, characterized in that, The multi-task discrimination module identifies configuration errors, including: Learn correct and incorrect configurations based on positive and negative samples, and train a classifier; The classifier is used to analyze the SCD generated file to generate a configuration error score; the configuration error score is represented as the weighted sum of the mean probability of error of all device parameters, the mean probability of error of topology, and the mean probability of error of device function in the SCD generated file output by the classifier.

6. The autonomous learning construction method for an engineering SCD verification model according to claim 1, characterized in that, The generator and discriminator are optimized through iterative training, including: The generator adjusts its loss function based on the topology accuracy score, text description fit score, and configuration error score output by the discriminator. The adjustment method is to add a quality score term based on the discriminator output to the generator's loss function on top of the adversarial loss term. The quality scoring items are the topological accuracy of the SCD-generated text, the text description fit, and the weighted sum of configuration errors.

7. The autonomous learning construction method for an engineering SCD verification model according to claim 6, characterized in that, In the calculation of the quality score items, topology accuracy is taken as 1 minus the topology accuracy score; text description fit is taken as 1 minus the text description fit score.

8. A self-learning construction system for an engineering SCD verification model, executing the self-learning construction method as described in any one of claims 1-7, characterized in that, The system includes: The sample data acquisition module is used for sample data collection and preprocessing; it extracts structured information from historical SCD configuration files and converts it into graph structure embedding representation; it collects historical unstructured text data and performs text embedding representation; the historical SCD configuration files include standard configuration files and non-standard configuration files, which are labeled as positive samples and negative samples, respectively. A generator design module is used to design a generator based on power grid requirements and standards to process positive sample data and text data. Two parallel processing networks are designed in the generator architecture to capture the internal graph structure features of the power grid in the graph structure embedding representation and the power grid equipment description features in the text embedding representation, respectively. The graph structure features and description features are fused to output SCD generated samples. Discriminator design module; used to design a discriminator based on the SCD structure; in the discriminator architecture, a multi-task discrimination module is designed to learn positive and negative samples and perform analysis on SCD-generated samples; including analyzing the accuracy of device topology in SCD-generated samples, judging the fit of text descriptions, and identifying configuration errors; The adversarial optimization module is used for iterative adversarial training and dynamic updates. It optimizes the generator and discriminator through iterative training. When a new SCD configuration file is input, the latest output discriminator is called for verification. The new SCD configuration file and verification results are incorporated into the collection of sample data to perform incremental learning.

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 computer program is loaded into the processor, it implements the autonomous learning construction method according to any one of claims 1-7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the autonomous learning construction method according to any one of claims 1-7.

Images