Combined extraction method for topological structure and electrical parameters of power grid

Abstract

The invention relates to the technical field of power system analysis and data processing, discloses a combined extraction method of a power grid topological structure and electrical parameters, and aims to solve the problems of fragmentation and inconsistency of topology and parameter information caused by data source splitting in traditional power grid design data processing. The method comprises the following steps: constructing a multi-modal analysis framework, and synchronously processing a wiring diagram, an equipment inventory, a specification and a nameplate image; extracting a power grid topological structure based on a graph neural network, identifying equipment symbols and reasoning complex connection logic; extracting electrical parameters such as voltage class and rated capacity from the text by using a named entity recognition and relation extraction technology; performing joint alignment and consistency verification on topology and parameters through cross-modal matching and power grid design specification constraint; and finally, a structured power grid digital model conforming to the IEC 61970 CIM standard is output, synchronous and high-precision combined extraction of the topological connection relation and the electrical attributes is realized, and high-quality data support is provided for power grid intelligent review and risk early warning.

Description

Technical Field

[0001] This invention relates to the field of power system analysis and data processing technology, specifically a method for the joint extraction of power grid topology and electrical parameters. Background Technology

[0002] With the continuous expansion of power grid infrastructure projects and the increasing complexity of their designs, the application of artificial intelligence technology in the design review process has become a key path to improve review efficiency and the scientific nature of decision-making. Currently, power grid design results are generally in the form of unstructured drawings, text reports, and scattered tables. Their core elements—power grid topology and electrical parameters—are often implicit in different file formats, lacking a unified and calculable digital representation. Traditional review processes heavily rely on manual review and cross-referencing, which is not only inefficient but also prone to overlooking critical issues such as topological logic errors and parameter mismatches due to information fragmentation or misunderstandings of standards, severely restricting review quality and investment accuracy. Especially in distribution network projects, with their complex equipment types, diverse wiring methods, and tightly coupled parameters, there is an urgent need for a technology that can automatically and jointly extract topological connection relationships and electrical attributes from heterogeneous design data, providing a structured data foundation for subsequent intelligent verification, scheme comparison, and risk warning.

[0003] The joint extraction of power grid topology and electrical parameters forms the underlying data engine for intelligent design review. This approach aims to overcome the limitations of traditional single-element identification by deeply integrating graphical semantic understanding and text information extraction to simultaneously reconstruct the physical connection logic of the power grid and equipment operating parameters, thus constructing a complete, consistent, and machine-readable digital model of the power grid. Its core challenge lies in achieving high-precision, strongly correlated extraction of topological relationships and parameter values ​​under real-world conditions such as variations in drawing symbols, ambiguity in text descriptions, and heterogeneity in data sources, ensuring strict alignment between the two at both spatial and logical levels.

[0004] Existing technologies mostly employ independent processing strategies: on the one hand, while computer vision-based drawing recognition methods can extract some equipment and connections, they struggle to accurately analyze complex wiring logic (such as ring networks and multiple power source access) and cannot correlate equipment nameplate parameters; on the other hand, natural language processing-based text extraction technologies can identify parameter values ​​but lack awareness of the equipment's role and connection relationships within the topology, resulting in information silos of "parameters without location" or "connections without attributes." Furthermore, existing methods generally ignore implicit constraints in power grid design specifications (such as matching rules between transformer capacity and outgoing line numbers), failing to perform consistency checks during extraction and leading to structural defects in the output model. Therefore, a joint extraction method is urgently needed that can collaboratively analyze graphics and text, integrate explicit connections and implicit rules, and ensure strict coupling between topology and parameters to support intelligent review and lean management throughout the entire power grid infrastructure project process. Summary of the Invention

[0005] The purpose of this invention is to provide a method for jointly extracting power grid topology and electrical parameters, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for jointly extracting power grid topology and electrical parameters includes the following specific steps: Step 1: Construct a multimodal power grid design data parsing framework, simultaneously access four heterogeneous data sources: primary wiring diagrams of the power grid, equipment and material lists, design specifications, and equipment nameplate images. Perform vectorization on the drawings, structured parsing on the text, and optical character recognition on the images to form a unified data input interface. Step 2: Perform topology extraction based on graph neural network, take the equipment symbols and connecting lines in the vectorized drawing as graph nodes and edges, classify the node types using a symbol recognition model, and parse complex topological logic such as busbars, feeders, ring networks and multiple power source access through connection relationship reasoning algorithm to generate an initial power grid topology map; Step 3: Perform electrical parameter extraction based on named entity recognition and relation extraction, identify equipment entities and their corresponding voltage levels, rated capacity, impedance parameters, protection settings and other attribute values ​​from the equipment material list and design specifications, and establish semantic association mapping between equipment entities and parameter values; Step 4: Perform topology-parameter joint alignment and consistency verification. Cross-modal matching is performed between the device nodes in the topology graph generated in Step 2 and the device entities extracted in Step 3. Precise association is established based on unique device identifiers, location descriptions, and type characteristics. Logical consistency verification is then performed on the joint results according to implicit constraint rules in power grid design specifications (such as the matching relationship between transformer capacity and the number of outgoing circuits), correcting conflicting terms. Among these, transformer capacity... Number of outgoing circuits The matching relationship can be represented as: in, This is the design redundancy factor (usually taken as 1.1~1.3). The reference capacity for a single circuit (e.g., 10 MVA) is used to constrain the rationality of the main transformer configuration and feeder scale. Step 5: Output the structured power grid digital model, encapsulate the verified topology connections and electrical parameters into an XML format file conforming to the IEC 61970 CIM standard, and support direct calling by subsequent intelligent review, scheme comparison and risk warning systems.

[0007] Preferably, in step 1, the vectorization processing employs a deep learning-based joint model for drawing segmentation and symbol detection. This model uses ResNet50 as the backbone network and combines it with an FPN feature pyramid structure to perform pixel-level segmentation of 12 core equipment symbols, including circuit breakers, disconnectors, transformers, busbars, and cables, in a primary wiring diagram. The symbol recognition accuracy is greater than or equal to 97%, and the connection line extraction completeness rate is greater than or equal to 95%. Preferably, in step 2, the symbol recognition model adopts a dual-branch convolutional neural network that integrates shape prior and contextual semantics. The shape branch extracts the geometric contour features of the equipment symbol, and the contextual branch captures its neighborhood layout information in the drawing. The two are weighted and integrated through a gating fusion mechanism, which improves the accuracy of distinguishing similar symbols (such as disconnect switches and grounding switches) to over 98.5%.

[0008] Preferably, the connection relationship reasoning algorithm in step 2 adopts a hybrid strategy based on graph traversal and rule guidance. First, a preliminary connection path is constructed through depth-first search. Then, based on the explicit rules in the power grid design specifications regarding bus segmentation, feeder branching, and protection coordination, ambiguous connections (such as loose connections and cross-connections) in the path are corrected. Finally, the topology logic accuracy is greater than or equal to 96%.

[0009] Preferably, in step 3, the named entity recognition model adopts the BERT-BiLSTM-CRF architecture and is pre-trained on the power grid domain corpus. It can recognize composite equipment entities including "110 kV main transformer", "35 kV outgoing line cabinet", and "zero-sequence current transformer". The entity recognition F1 value is greater than or equal to 95%. The parameter value extraction adopts a method based on regular template and numerical context joint matching. Non-standard expressions (such as "about 10 MVA" and "not greater than 500 A") are normalized, and the parameter normalization accuracy is greater than or equal to 94%.

[0010] Preferably, in step 4, the cross-modal matching adopts a multi-granularity feature alignment mechanism, with the matching of unique device identifiers as the first priority. When an identifier is missing, the location description similarity (based on cosine distance calculation) and type feature consistency (based on device classification confidence) are used sequentially for matching, with a matching success rate of over 98%. The implicit constraint rule base contains more than 200 typical power grid design specifications, covering key coupling relationships such as main transformer-outgoing line, line-protection, and reactive power compensation-load, with a verification coverage rate of over 90%.

[0011] Preferably, in step 4, the logical consistency verification adopts a dual verification mechanism based on a rule engine and a graph neural network. The rule engine performs hard constraint checks (such as capacity overrun and protection deficiency), while the graph neural network learns soft patterns in historical compliant designs and provides early warnings for potential structural defects (such as insufficient topological redundancy and abnormal parameter gradients). The combined accuracy of the dual verification is greater than or equal to 97%.

[0012] Preferably, the structured power grid digital model in step 5 includes a complete list of topology nodes, a list of edge connections, and a node attribute table. The node attribute table fields include 15 core parameters such as device ID, device type, voltage level, rated capacity, short-circuit impedance, and protection configuration. The model file size is compressed to less than 5% of the original design data, and the loading response time is less than 1 second.

[0013] Preferably, the method further includes an incremental learning and feedback optimization module. When a new type of equipment symbol or a new wiring method appears, the system automatically collects expert correction samples and updates the symbol recognition model and connection relationship reasoning algorithm through an online fine-tuning mechanism. The model update cycle is less than 24 hours, ensuring the system's continuous adaptability to design evolution.

[0014] Preferably, the multimodal power grid design data parsing framework supports automatic recognition and conversion of four mainstream formats: PDF, DWG, DOCX, and JPG, with 100% format compatibility coverage. It requires scanned drawings to have a resolution of at least 300 dpi and achieves an automatic character encoding recognition accuracy of at least 99% for text files. Compared with the prior art, the beneficial effects achieved by the present invention are: 1. This invention constructs a multimodal power grid design data parsing framework, which realizes unified access and structured preprocessing of heterogeneous design data such as drawings, texts, and images, solving the information fragmentation problem caused by the fragmentation of data sources in traditional methods; through the coordinated execution of topology extraction based on graph neural networks and electrical parameter extraction based on named entity recognition, it breaks through the technical limitations of independent processing of single elements, and for the first time realizes the synchronous and joint extraction of topological connection relationships and electrical attributes.

[0015] 2. By introducing a topology-parameter joint alignment and consistency verification mechanism, and deeply integrating the unique identifier of equipment, location semantics, and implicit constraints in power grid design specifications (including quantitative matching relationships such as the main transformer capacity and the number of outgoing circuits), the output model is ensured to have strict consistency in spatial location and logical rules, effectively avoiding information silos such as "parameters without location" or "connections without attributes". The final structured power grid digital model can directly serve upper-level applications such as intelligent review, scheme comparison and risk warning, significantly improving the automation level, decision-making scientificity and management leanness of power grid infrastructure engineering design review, and providing a solid data foundation for the high-quality development of the power grid. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method for jointly extracting power grid topology and electrical parameters proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the collaborative execution of topology extraction based on graph neural network and electrical parameter extraction based on named entity recognition in this invention; Figure 3 This is a logical flowchart of the multimodal power grid design data parsing framework in this invention; Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow of topology-parameter joint alignment and consistency verification in this invention; Figure 5 This is a logical flow diagram of the integration of the structured power grid digital model output and upper-layer applications in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0019] Currently, with the continuous expansion of power grid infrastructure projects and the increasing complexity of their designs, the application of artificial intelligence technology in the design review process has become a key path to improve review efficiency and the scientific nature of decision-making. Current power grid design results are generally in the form of unstructured drawings, text reports, and scattered tables. Their core elements—power grid topology and electrical parameters—are often implicit in different file formats, lacking a unified and calculable digital representation. Traditional review processes heavily rely on manual review and cross-referencing, which is not only inefficient but also prone to overlooking critical issues such as topological logic errors and parameter mismatches due to information fragmentation or misunderstandings of standards, severely restricting review quality and investment accuracy. Especially in distribution network projects, with their complex equipment types, diverse wiring methods, and tightly coupled parameters, there is an urgent need for a technology that can automatically and jointly extract topological connection relationships and electrical attributes from heterogeneous design data, providing a structured data foundation for subsequent intelligent verification, scheme comparison, and risk warning. To address the aforementioned technical issues, this invention proposes a multimodal power grid design data parsing framework that collaboratively performs topology extraction based on graph neural networks and electrical parameter extraction based on named entity recognition. It also introduces a joint alignment and consistency verification mechanism for topology and parameters, deeply integrating unique equipment identifiers, location semantics, and implicit constraints in power grid design specifications. This ensures strict consistency of the output model at both the spatial location and logical rule levels, and is applied to a joint extraction method for power grid topology and electrical parameters.

[0020] refer to Figure 1 The overall technical architecture of this invention includes a multimodal power grid design data parsing framework, a topology extraction module based on graph neural networks, an electrical parameter extraction module based on named entity recognition, a topology-parameter joint alignment and consistency verification module, and a structured power grid digital model output module. These modules work collaboratively to achieve end-to-end conversion from heterogeneous design data to a structured, machine-readable power grid digital model.

[0021] In the aforementioned method for jointly extracting power grid topology and electrical parameters, step 1 involves constructing a multimodal power grid design data parsing framework. This framework simultaneously accesses four heterogeneous data sources: primary wiring diagrams, equipment and material lists, design specifications, and equipment nameplate images. Drawings are vectorized, text is structured, and images are optical character recognition (OCR) to form a unified data input interface. Specifically, the multimodal power grid design data parsing framework supports automatic recognition and conversion of four mainstream formats: PDF, DWG, DOCX, and JPG, achieving 100% format compatibility coverage. For primary wiring diagrams, regardless of whether the original format is vector DWG or bitmap PDF / JPG, preprocessing is performed first. For bitmap formats, a scanning resolution of at least 300 dpi is required to ensure the clarity of symbol details. Subsequently, a deep learning-based joint model for drawing segmentation and symbol detection is used for vectorization. This model uses ResNet50 as the backbone network and combines it with an FPN feature pyramid structure to effectively fuse feature information at different scales. The model performs pixel-level segmentation of 12 core equipment symbols, including circuit breakers, disconnectors, transformers, busbars, and cables, in primary wiring diagrams, achieving a symbol recognition accuracy of ≥97%. Simultaneously, the model can accurately extract connecting lines, including straight lines, broken lines, and arcs, with a connection line extraction completeness rate of ≥95%. For text documents such as equipment material lists and design specifications, the system first automatically identifies their character encoding (e.g., UTF-8, GBK), achieving an automatic recognition accuracy of ≥99%, and then performs structured parsing, converting tables, paragraphs, headings, and other elements into structured text with semantic tags. For equipment nameplate images, high-precision optical character recognition technology is used to extract text information from the nameplates, including equipment model, serial number, and rated parameters. After the above processing, all four types of data sources are converted into a unified intermediate data format, forming a standardized data input interface, providing a high-quality, structured data foundation for subsequent joint extraction.

[0022] refer to Figure 3 This diagram details the logical flow of the multimodal power grid design data analysis framework. This framework serves as the entry point for the entire method, and its output directly supplies the two core modules: topology extraction and electrical parameter extraction.

[0023] In the aforementioned method for jointly extracting power grid topology and electrical parameters, step 2 involves performing topology extraction based on a graph neural network. Equipment symbols and connecting lines in the vectorized drawing are used as graph nodes and edges. A symbol recognition model is used to classify node types, and a connection relationship reasoning algorithm is used to parse complex topological logic such as buses, feeders, ring networks, and multiple power source access, generating an initial power grid topology map. Specifically, the symbol recognition model in step 2 employs a dual-branch convolutional neural network that integrates shape priors and contextual semantics. The shape branch focuses on extracting the geometric contour features of equipment symbols, such as aspect ratio, closure, and internal structure; the context branch captures the neighborhood layout information of the equipment symbol in the drawing, such as whether there are buses around it, its relative distance and orientation to other equipment, etc. The feature vectors of these two branches are weighted and integrated through a gating fusion mechanism that dynamically adjusts the contribution of the two branches according to the characteristics of the input symbols. This design significantly improves the ability to distinguish similar symbols (such as disconnectors and grounding switches), increasing the accuracy to over 98.5%. After classifying all nodes, the system uses device symbols as graph nodes and connecting lines as graph edges to construct an initial graph structure. Subsequently, a connection relationship reasoning algorithm is introduced, employing a hybrid strategy based on graph traversal and rule guidance. First, the entire graph structure is traversed using depth-first search or breadth-first search to construct preliminary connection paths and identify basic units such as buses and feeders. However, ambiguities such as loose connections and incorrect cross-connections often exist in the drawings, which cannot be accurately resolved by graph traversal alone. Therefore, the algorithm further modifies the preliminary paths based on explicit rules in power grid design specifications regarding bus segmentation, feeder branching, and protection coordination. For example, if a rule stipulates that electrically isolated devices should not exist on the same bus segment, the algorithm will merge or split bus nodes accordingly; if a rule stipulates that feeders must originate from buses, the algorithm will correct devices incorrectly connected to other feeders. Through this hybrid strategy, the final generated initial power grid topology diagram is logically highly accurate, with a topology logic accuracy rate greater than or equal to 96%.

[0024] refer to Figure 2 This figure illustrates the core principle framework of the collaborative execution of topology extraction based on graph neural networks and electrical parameter extraction based on named entity recognition. The output of the topology extraction module, i.e., the initial power grid topology diagram, will be jointly aligned with the output of the electrical parameter extraction module.

[0025] In the aforementioned method for jointly extracting power grid topology and electrical parameters, step 3 involves extracting electrical parameters based on named entity recognition and relation extraction. This process identifies equipment entities and their corresponding voltage levels, rated capacities, impedance parameters, protection settings, and other attribute values ​​from equipment material lists and design specifications, and establishes a semantic association mapping between equipment entities and parameter values. Specifically, the named entity recognition model in step 3 adopts a BERT-BiLSTM-CRF architecture. This model is first pre-trained on a large-scale general corpus, and then undergoes domain-adaptive fine-tuning on specially collected power grid domain corpora (including historical design specifications, equipment manuals, and regulatory documents) to deeply understand the professional terminology and expression habits of the power grid domain. The model can accurately identify composite equipment entities, including "110 kV main transformer," "35 kV outgoing line cabinet," and "zero-sequence current transformer," with an entity recognition F1 score greater than or equal to 95%. After identifying the equipment entities, the relation extraction module is responsible for extracting their corresponding attribute values ​​from the surrounding text context. Parameter value extraction employs a method based on joint matching of regular templates and numerical context. The system predefines a series of regular expression templates for matching standard parameter representations (such as "rated capacity: 50 MVA"). For non-standard representations (such as "approximately 10 MVA" or "not greater than 500 A"), the system normalizes them to standard numerical ranges or point estimates by analyzing modifiers in the numerical context (such as "approximately", "not greater than", "typical value", etc.). For example, "approximately 10 MVA" is normalized to "10.0 MVA", and "not greater than 500 A" is normalized to "<=500 A". This normalization process ensures the standardization and computability of parameter values, with a parameter normalization accuracy greater than or equal to 94%. Ultimately, each identified device entity establishes a precise semantic association mapping with its corresponding multiple electrical parameters, forming a structured device-parameter knowledge base.

[0026] In the aforementioned method for jointly extracting power grid topology and electrical parameters, step 4 involves implementing joint alignment and consistency verification of the topology and parameters. This involves cross-modal matching of the device nodes in the topology graph generated in step 2 with the device entities extracted in step 3. Precise association is performed based on the device's unique identifier, location description, and type characteristics. Logical consistency verification is then performed on the joint results according to implicit constraint rules in power grid design specifications, correcting any conflicts. Specifically, cross-modal matching in step 4 is the core of the entire method, employing a multi-granularity feature alignment mechanism. The first priority for matching is the device's unique identifier, such as the device's number in the material list or the serial number on its nameplate. If both the topology node and the device entity contain this identifier and are consistent, the matching is completed directly. When the identifier is missing or inconsistent (which is common in actual engineering), the system sequentially uses suboptimal matching strategies. The second priority is location description similarity. The system converts the coordinates of the topology node in the drawing into a text description (e.g., "second bay on the east side of the 110 kV I section busbar") and compares it with the location description of the equipment entity in the text (e.g., "installed in the second bay on the east side of the 110 kV I section busbar"). The system calculates the cosine distance between the two based on word embeddings. If the distance is less than a preset threshold, the match is considered successful. The third priority is type feature consistency. The system compares the type with the highest confidence in symbol recognition of the topology node with the type of the equipment entity. If they match and both have confidence scores higher than the threshold, the match is complete. Through this multi-granularity, prioritized matching strategy, the overall matching success rate reaches over 98%. After matching is completed, the consistency verification stage begins. The verification is based on an implicit constraint rule library containing over 200 typical power grid design specifications. This library covers key coupling relationships such as transformer-outgoing line, line-protection, and reactive power compensation-load, with a verification coverage rate of over 90%. The verification adopts a dual verification mechanism based on a rule engine and a graph neural network. The rules engine is responsible for performing hard constraint checks, such as checking whether the main transformer capacity is exceeded or whether critical equipment lacks protection configurations. One key quantitative rule is: main transformer capacity, main transformer capacity, main transformer capacity. Number of outgoing circuits The matching relationship can be represented as: in, This is the design redundancy factor (usually taken as 1.1~1.3). A reference capacity of 10 MVA is typically used for a single loop. The rule engine automatically calculates this inequality, marking any violations as conflict terms. The graph neural network is responsible for learning soft patterns in historical compliant designs. It is trained to identify patterns that, while not violating hard rules, pose potential risks, such as insufficient topology redundancy (e.g., critical loads powered by a single power source) and abnormal parameter gradients (e.g., excessive impedance differences between adjacent lines). It also provides warnings about these potential structural defects. For marked conflict terms and warning terms, the system generates correction suggestions or performs corrections directly in the automated process (e.g., inferring missing protection devices based on rules and adding them to the topology). The dual-verification combined accuracy is greater than or equal to 97%, ensuring high consistency and reliability of the final joint result.

[0027] refer to Figure 4 This diagram clearly depicts the multi-level interactions and data flow between topology-parameter joint alignment and consistency verification. Data from the topology extraction and parameter extraction modules converge here, undergo matching and double verification, and output a high-quality joint model.

[0028] In the aforementioned method for jointly extracting power grid topology and electrical parameters, step 5 outputs a structured power grid digital model. The verified topology connections and electrical parameters are encapsulated into an XML file conforming to the IEC 61970 CIM standard, supporting direct access by subsequent intelligent review, scheme comparison, and risk warning systems. Specifically, the structured power grid digital model in step 5 comprises three core components: a complete list of topology nodes, an edge connection list, and a node attribute table. The topology node list records the unique IDs and types of all devices in the power grid; the edge connection list records the electrical connections between all devices; and the node attribute table appends detailed electrical parameters to each device node, including 15 core parameters such as device ID, device type, voltage level, rated capacity, short-circuit impedance, and protection configuration. All this information is modeled and serialized according to the IEC 61970 CIM standard and encapsulated into a single XML file. This standard is an internationally recognized standard for power system information models, ensuring model interoperability and broad system compatibility. Through efficient compression algorithms, the model file size is compressed to less than 5% of the total size of the original design data, significantly saving storage and transmission costs. Meanwhile, due to the model's highly structured nature, it loads and parses in memory extremely quickly, with a loading response time of less than 1 second, which can meet the real-time requirements of upper-layer applications.

[0029] Furthermore, the method also includes an incremental learning and feedback optimization module. In practical applications, power grid design technology is constantly evolving, and new types of equipment symbols or novel wiring methods may emerge, exceeding the recognition range of the current model. Therefore, the system is designed with a closed-loop feedback optimization mechanism. When experts discover errors in model extraction and manually correct them during use, these corrected samples are automatically collected and marked. The incremental learning module is triggered periodically (or after accumulating a certain number of samples) to fine-tune the symbol recognition model and connection relationship inference algorithm online. This fine-tuning requires only a small number of samples and a short time; the model update cycle is less than 24 hours, ensuring the system's continuous adaptability to new trends in power grid design and its long-term effectiveness.

[0030] refer to Figure 5 The diagram illustrates the integration logic between the structured power grid digital model output and upper-level applications (such as intelligent review, scheme comparison and selection, and risk warning system), demonstrating the core value of this invention as a bottom-level data engine.

[0031] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

[0032] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for jointly extracting power grid topology and electrical parameters, characterized in that: The specific steps include the following: Step 1: Construct a multimodal power grid design data parsing framework, simultaneously access four heterogeneous data sources: primary wiring diagrams of the power grid, equipment and material lists, design specifications, and equipment nameplate images. Perform vectorization on the drawings, structured parsing on the text, and optical character recognition on the images to form a unified data input interface. Step 2: Perform topology extraction based on graph neural network, take the equipment symbols and connecting lines in the vectorized drawing as graph nodes and edges, classify the node types using a symbol recognition model, and parse complex topological logic such as busbars, feeders, ring networks and multiple power source access through connection relationship reasoning algorithm to generate an initial power grid topology map; Step 3: Perform electrical parameter extraction based on named entity recognition and relation extraction, identify equipment entities and their corresponding voltage levels, rated capacity, impedance parameters, protection settings and other attribute values ​​from the equipment material list and design specifications, and establish semantic association mapping between equipment entities and parameter values; Step 4: Perform topology-parameter joint alignment and consistency verification. Perform cross-modal matching between the device nodes in the topology graph generated in Step 2 and the device entities extracted in Step 3. Make precise associations based on the device's unique identifier, location description and type characteristics. Perform logical consistency verification on the joint results according to the implicit constraint rules in the power grid design specifications and correct conflict items. Step 5: Output the structured power grid digital model, encapsulate the verified topology connections and electrical parameters into an XML format file conforming to the IEC 61970 CIM standard, and support direct calling by subsequent intelligent review, scheme comparison and risk warning systems.

2. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 1, the vectorization process adopts a deep learning-based drawing segmentation and symbol detection joint model. This model uses ResNet50 as the backbone network and combines it with the FPN feature pyramid structure to perform pixel-level segmentation of 12 types of core equipment symbols such as circuit breakers, disconnect switches, transformers, busbars and cables in a primary wiring diagram. The symbol recognition accuracy is greater than or equal to 97%, and the connection line extraction completeness rate is greater than or equal to 95%.

3. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 2, the symbol recognition model adopts a dual-branch convolutional neural network that integrates shape prior and contextual semantics. The shape branch extracts the geometric contour features of the device symbol, and the contextual branch captures its neighborhood layout information in the drawing. The two are weighted and integrated through a gating fusion mechanism, and the accuracy of distinguishing similar symbols is greater than or equal to 98.5%.

4. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 2, the connection reasoning algorithm adopts a hybrid strategy based on graph traversal and rule guidance. First, a preliminary connection path is constructed through depth-first search. Then, based on the explicit rules in the power grid design specifications regarding bus segmentation, feeder branching, and protection coordination, ambiguous connections in the path are corrected. Finally, the topology logic accuracy is greater than or equal to 96%.

5. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 3, the named entity recognition model adopts the BERT-BiLSTM-CRF architecture and is pre-trained on the power grid domain corpus. It can recognize composite equipment entities with an entity recognition F1 score of greater than or equal to 95%. The parameter value extraction adopts a method based on regular template and numerical context joint matching to normalize non-standard expressions, and the parameter normalization accuracy is greater than or equal to 94%.

6. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 4, cross-modal matching adopts a multi-granularity feature alignment mechanism, with the matching of unique device identifiers as the first priority. When an identifier is missing, the similarity of location description and the consistency of type features are used for matching in sequence, with a matching success rate of over 98%. The implicit constraint rule base contains more than 200 typical power grid design specifications, covering key coupling relationships such as main transformer-outgoing line, line-protection, and reactive power compensation-load, with a verification coverage rate of over 90%.

7. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 4, the logical consistency verification adopts a dual verification mechanism based on a rule engine and a graph neural network. The rule engine performs hard constraint checks, while the graph neural network learns soft patterns in historical compliance designs to provide early warnings of potential structural defects. The combined accuracy of the dual verification is greater than or equal to 97%.

8. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: In step 5, the structured power grid digital model includes a complete list of topological nodes, a list of edge connections, and a node attribute table. The node attribute table includes 15 core parameters such as device ID, device type, voltage level, rated capacity, short-circuit impedance, and protection configuration. The model file size is compressed to less than 5% of the original design data, and the loading response time is less than 1 second.

9. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: It also includes an incremental learning and feedback optimization module. When new types of equipment symbols or new wiring methods appear, the system automatically collects expert correction samples and updates the symbol recognition model and connection relationship reasoning algorithm through an online fine-tuning mechanism. The model update cycle is less than 24 hours.

10. The method for jointly extracting power grid topology and electrical parameters according to claim 1, characterized in that: The multimodal power grid design data parsing framework supports automatic recognition and conversion of four mainstream formats: PDF, DWG, DOCX, and JPG, with 100% format compatibility coverage. It requires scanned drawings to have a resolution of no less than 300 dpi and has an automatic character encoding recognition accuracy of greater than or equal to 99% for text files.

Images