Neural Network Recognition Method for Substation Based on Electrical Industry Knowledge Graph

Abstract

The present invention discloses a method for identifying a substation neural network based on an electrical industry knowledge graph. First, based on the design standards of substations and distribution stations, a knowledge graph of substation design engineering is constructed; then, based on the deep learning of recurrent neural networks, the parameters of electrical system diagrams are intelligently identified; the parameters of electrical system diagrams are learned as the electronic parameters of entities and the relationships between them through graph neural networks, and in combination with the knowledge graph of substation design engineering, the positions, quantities, and relevant electrical parameters of electrical equipment are matched with the standards, and a floor plan is automatically generated based on the matching information; finally, based on a generative adversarial network, the floor plan is mapped to a 3D model diagram, and the building structure model of the substation and distribution station is modeled through a 3D platform to generate a 3D model diagram of the substation and distribution station equipment; thereby realizing automatic drawing, improving the design efficiency, greatly reducing the burden on designers, and at the same time avoiding misjudgment or omission of key parameters caused by human factors.

Description

Technical Field

[0001] The present invention relates to the field of substation and distribution station design and the field of artificial intelligence, and particularly relates to a method for identifying a neural network of a substation and distribution station based on an electrical industry knowledge graph. Background Art

[0002] With the continuous development of the power industry, the design of substations and distribution stations and their internal electrical equipment has become increasingly complex. Therefore, during the design process of substations and distribution stations for complex users, a large amount of manpower and material resources are consumed, and there are problems of insufficient refinement and intelligence. Moreover, there are still inevitable situations of parameter misjudgment or omission caused by humans to varying degrees.

[0003] As an intelligent and efficient knowledge organization method, the knowledge graph has been widely used in general fields and specific fields such as finance, healthcare, and geography. Since the data in the power field has the same characteristics as that in other fields, such as large scale, diverse sources, and inconsistent data structures, many application scenarios and ideas of the knowledge graph can be extended to the power field. By constructing a knowledge graph for the power field, the scattered and diverse-structured data in the power field can be effectively processed and organized to ensure that the data of the power grid can be entered at one place and used throughout the network, providing guarantees for the authenticity, consistency, and integrity of information, and achieving the goal of building the national power grid into an integrated enterprise-level information integration platform.

[0004] Therefore, by designing a knowledge graph modeling method for the design engineering field of substations and distribution stations for complex users, using machine learning technology to construct the information of the schema layer and data layer of the knowledge graph, and using an inference engine for intelligent identification and judgment, the problem of low current design efficiency can be solved. Summary of the Invention

[0005] To solve the deficiencies of the prior art, achieve automatic drawing, greatly reduce the burden on designers, improve the automatic configuration ability of the system, and also avoid misjudgment or omission of key parameters caused by human factors, the present invention adopts the following technical solutions:

[0006] A method for identifying a neural network of a substation and distribution station based on an electrical industry knowledge graph includes the following steps:

[0007] Step S1: Based on the design standards of substations and distribution stations, construct a knowledge graph for substation design engineering;

[0008] Step S2: Based on the deep learning of a recurrent neural network, intelligently identify the parameters of the electrical system diagram;

[0009] Step S3: Use a graph neural network to learn the electrical system diagram parameters as the electronic parameters of entities and the relationships between them. Combine with the substation design engineering knowledge graph to match the location, quantity, and relevant electrical parameters of electrical equipment with the standard, and automatically generate a floor plan based on the matching information;

[0010] Step S4: Based on a generative adversarial network, map the floor plan to a 3D model diagram, and complete the modeling of the substation building structure model through a 3D platform to generate a 3D model diagram of the substation equipment.

[0011] Further, in step S2, the electrical system diagram parameters include room information and primary trunk line electrical parameters. The intelligent recognition process includes the following steps:

[0012] Step S2.1: Obtain the electrical system diagram and perform image segmentation. Each segmented small diagram contains one type of electrical parameter;

[0013] Step S2.2: Mark the electrical parameter x1 in the small diagram with the correct category label. Mark parameter x1 as label y1 (for example: label y1 is the loop usage parameter), parameter x2 as y2 (for example: label y2 is the load parameter), and use the labeled electrical parameters as the training set and input them into the neural network VGG16;

[0014] Step 2.3: The training set (such as parameter x1) is input into the 3*3*C convolutional layer Conv5 of the neural network VGG16, where 3*3 represents the size of the convolutional window and C represents the number of channels. Perform feature extraction of the electrical parameters through convolutional operations to obtain the feature map x2 output by the convolutional layer;

[0015] Step 2.4: The feature map x2 is input into a bidirectional long short-term memory layer. Each row of the feature map x2 is cyclically connected to the bidirectional long short-term memory layer to obtain a 256-dimensional Bi-LSTM, and finally the output feature map x3 is obtained;

[0016] Step 2.5: The feature map x3 is input into an encoder composed of 5 convolutional layers for encoding operations to obtain the encoded output feature x4. Perform multiplication and addition operations on the encoded output feature in sequence, and then sample through the normal distribution N(0,1) to obtain the feature x6. The feature x6 is decoded by the decoder to obtain the feature x7, and then through the fully connected layer, the room coordinate position information, category score, and side compensation are obtained respectively. Based on the side compensation, the side improvement offset of the room coordinates is obtained to improve the positioning accuracy, and the predicted electrical parameter category Y is obtained through the category score;

[0017] Step 2.6: Compare the predicted electrical parameter category Y with the correctly marked category, calculate the loss based on the loss function, and train the neural network VGG16;

[0018] Step 2.7: Input the electrical system diagram to be predicted into the trained neural network VGG16 to obtain the predicted electrical parameter categories and room coordinate position information.

[0019] Furthermore, the generated knowledge graph is used for electrical system diagram parameter recognition. Based on the knowledge graph where entities and the relationships between entities are based on design standards, the knowledge graph weights are generated, and a recursive neural network deep learning is introduced to improve the accuracy of electrical system diagram parameter recognition and avoid misidentifying the electrical parameters as other similar ones.

[0020] Furthermore, in step S3, the combination of the electronic parameters as entities and the relationships between them with the substation design engineering knowledge graph includes the following steps:

[0021] Step S3.1: Map the entity space to the relationship space to make the distinguishability of relationships stronger. When two entities have similar meanings, the distance is close; when they do not have similar meanings, the distance is far. Define the mapping operator Map and use two methods, Word2vec and the skip-gram model, to project from the entity space to the embedded relationship space.

[0022] Step S3.2: Add the parameter b of the activation function with a non-linear transformation ijt , which is the same as the graph attention network operation. Calculate the correlation model weight coefficient using adjacent entity nodes (triples) on the knowledge graph:

[0023] a ijt = softmax jt (b ijt )

[0024] softmax jk represents the activation function of the j-th entity based on the relationship t, where t represents the relationship type, and i, j represent the entity types. After adding the relationship of the edge using the graph neural network and performing multiple iterations on the entities, the information and relationships between the entity nodes and their n-th order neighbor entity nodes are obtained, so that all entities at different distances on the knowledge graph can more efficiently and globally share the information they carry through the edges to assist in completing prediction and reasoning tasks, such as for a large number of geometric constraints (non-overlapping, adjacent arrangement, parallel arrangement, etc.) and engineering constraints (minimum distance from the wall, minimum distance between equipment, minimum width of the pedestrian passage, etc.) in the plane layout planning.

[0025] Furthermore, in step S3.1, each relationship triple is concatenated by a connection method and then multiplied by the linear transformation parameter matrix W 1 to be transformed into the relationship space:

[0026]

[0027] where W 1 is learned by the Skip-gram method of the skip-gram model. t represents the relationship type, and i and j represent entity types, and i≠j. An entity may have information at multiple levels, and different relationships may focus on different levels of the entity. The optimal dimensions for expressing entities and relationships are not necessarily the same. Therefore, mapping them to the same space may limit the model's performance. Therefore, entity pairs with the same relationship are clustered.

[0028] Furthermore, a variational autoencoder generative adversarial network is adopted, including a generator G and a discriminator D. The two-dimensional planar layout drawing x is used as the input. The latent vector z is obtained through the variational autoencoder and sent into the generator G. The generator G randomly samples a multi-dimensional latent vector z from a probabilistic latent space and maps it into a 64*64*64 cube, which is represented by the three-dimensional volume space G(z). The discriminator D outputs the confidence D(x) based on the input of the generator G to determine whether it matches the input two-dimensional planar layout drawing x. The binary cross-entropy is used as the classification loss to train the adversarial network. The loss function of the model is jointly determined by the reconstruction error, the cross-entropy of the adversarial network, and the KL distance. The three-dimensional substation building structure model is modeled through a three-dimensional platform, mainly including buildings, upper steel frame structures, lower pile foundations, and main transformer equipment foundations, etc. The attributes of building walls, doors, windows, etc. are assigned strictly in accordance with the modeling specifications.

[0029] Furthermore, based on the generated three-dimensional model drawing, collision detection is carried out. Electrical equipment with intersections in the three-dimensional drawing is screened. At the same time, through the identified electrical system drawing parameters and the knowledge graph constructed based on the substation design standard, entity information and entity relationships of room information and electrical parameters are obtained. The installation and deployment requirements in the substation design process are integrated to supplement the screening of collisions of electrical equipment and adjust the collision situations. Since the knowledge graph is constructed based on the substation design standard, and the substation design standard includes the requirements for avoiding collisions when installing and deploying equipment of different sizes in different room structures, it is necessary to supplement and screen the collision detection by combining the knowledge graph and electrical system drawing parameters. Then, the collision detection results are located and given, and equipment openings are reserved at the collision positions. Through collision detection, collision problems in the project can be screened out, and design modifications and optimizations can be carried out according to the collision inspection results to reduce collision problems between buildings, pipelines, and electrical equipment. The collision detection of the three-dimensional modeling platform can eliminate collision problems such as pipeline conflicts before construction, minimizing design errors that may be encountered during the construction stage and reducing construction delays and capital waste caused by collision problems.

[0030] Further, in step S1, according to relevant design standards of the substation, such as national standards and industry standards, a rule library is established, entities of the substation in the rule library and the relationships between the entities based on the design standards are obtained, and a knowledge graph is constructed; in a knowledge fusion manner, from the perspective of similarity, it is determined whether two entities are the same entity or can be merged into the same entity object classification; the judgment based on similarity is to judge the similarity of the character descriptions, attributes / attribute values, and semantic structures of the two entities. The greater the similarity, the greater the possibility that the two entities are the same entity, or the more likely these two entities point to the same entity object.

[0031] Further, in step S1, the knowledge graph is represented by a triple K(S, O, R), where S represents the subject, O represents the object, and the relationship is defined as R:={SR|Value}. The relationship has a value attribute and is irreversible, that is, the subject S and the object O cannot be reversed.

[0032] The advantages and beneficial effects of the present invention are as follows:

[0033] The substation neural network recognition method based on the electrical industry knowledge graph of the present invention makes the graphic elements and data models correspond one by one. On this basis, the power drawing standard is introduced to construct a knowledge graph for the complex user substation design engineering field. On the basis of the knowledge graph, through the intelligent recognition method of electrical system diagram parameters based on the recursive neural network learning algorithm, and at the same time the method for generating the substation equipment layout plan based on graph reasoning, the relevant rules are defined through graph reasoning to automatically generate various types of power design drawings, realize the automatic understanding of electrical system drawings and building plan drawings, and realize the automatic generation of substation design schemes on the basis of automatically obtaining design parameters, and realize the rapid generation of including equipment layout plans, civil construction drawings, grounding lighting drawings, etc. It lays a foundation for realizing electrical constraint optimization and environmental factor strategy control, and realizing the personalized optimization strategy of microgrid substation design for thousands of people. Finally, 80% of the workload of the original traditional design is automatically completed in the AI program design, minimizing human intervention as much as possible, solving the problem of automatic generation of substation design schemes in medium and low voltage substations, and solving the problems of long design cycle, large amount of manual intervention workload, and difficulty in checking the correctness of traditional manual design methods.

[0034] Description of the drawings

[0035] Figure 1 is the flowchart of the substation neural network recognition method based on the electrical industry knowledge graph in the embodiment of the present invention.

[0036] Figure 2 is a partial knowledge graph schematic diagram of the switch entity in the substation electrical field in the embodiment of the present invention. Detailed implementation manners

[0037] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only for the purpose of illustrating and explaining the present invention, and are not intended to limit the present invention.

[0038] As Figure 1 shown, a method for identifying a substation neural network based on an electrical industry knowledge graph includes the following steps:

[0039] Step S1: Based on the design standards of the substation, construct a knowledge graph of the substation design project through a graph neural network;

[0040] According to the relevant design standards of the substation, such as national standards and industry standards, establish a rule base, obtain the entities of the substation in the rule base and the relationships between the entities based on the design standards, and construct a knowledge graph;

[0041] The design engineering problems of complex user substations belong to the vertical industry field. The construction of its domain knowledge graph is respectively composed of substation design constraint knowledge representation, environmental knowledge acquisition, knowledge storage and knowledge fusion, which constitute a cyclic iterative life cycle. The key technologies for constructing a knowledge graph include in-domain knowledge representation modeling, entity recognition and entity linking, relationship event extraction, discovery of implicit relationships, etc., including the discovery and restoration of missing information in the knowledge graph.

[0042] Knowledge fusion is to judge whether two entities are the same entity or can be merged into the same entity object classification from the perspective of similarity. Specifically, judge the similarity of the character descriptions, attributes / attribute values and semantic structures of the two entities. The greater the similarity, the greater the possibility that the two entities are the same entity, or the more likely these two entities point to the same entity object. As Figure 2 shown, the possible relationships between entities are represented by connecting lines.

[0043] The knowledge graph is represented as a triple K(S, O, R), where S represents the subject, O represents the object, and the relationship is defined as R:={SR|Value}. The relationship has a value attribute and is irreversible, that is, the subject S and the object O cannot be reversed. Specific semantic relationships in the electrical field are exemplified as follows:

[0044] Clause [4.1.6]:... In the low-voltage distribution device, an appropriate number of spare circuits should be left.

[0045] K_50053_416={S_50053_416={low-voltage distribution device}, O_50053_416={spare circuit},

[0046] R_50053_210={retain|appropriate number}}

[0047] Clause [6.2.6]: For a power distribution room with a length greater than 7m, two safety exits shall be provided and preferably arranged at both ends of the power distribution room.

[0048] K_50053_6261 = {S_50053_416 = {power distribution room}, O_50053_416 = {length},

[0049] R_50053_211 = {greater than | 7m}}

[0050] K_50053_6262 = {S_50053_416 = {power distribution room}, O_50053_417 = {safety exit}, R_50053_210 = {placed | 2}}

[0051] Regulation: K_50053_6261 → K_50053_6262

[0052] Step S2: Based on deep learning of recurrent neural networks, intelligently identify the parameters of the electrical system diagram;

[0053] Through the method of identifying the parameters of the substation, quickly and reliably identify the information of the room where the substation is located (such as the outline of the outer wall, the position of doors and windows, and obstacles such as internal columns) and the electrical parameters of the primary trunk line (circuit use, load level, power supply range, installation capacity, demand factor, power factor, distribution box parameters, etc.) from the building floor plan provided by the user;

[0054] Based on the convolutional neural network (CNN) and the recurrent neural network (RNN), combine CNN, RNN and the encoding - decoding unit. By introducing the ideas of "differential" and "variational inference", use a recurrent neural network for proposed information variational inference to narrow the difference between the training sample distribution and the real image data distribution, so as to achieve the purpose of training the common and different features of the images of electrical equipment and improve the accuracy of identifying the parameters of the electrical system diagram. The specific steps are as follows:

[0055] Step S2.1: Obtain the electrical system diagram and perform image segmentation. Each small segmented image contains one kind of electrical parameter;

[0056] Step S2.2: Mark the electrical parameters in the small image with the correct category. Mark the parameter x1 as the label y1 (for example: the label y1 is the circuit use parameter), and the parameter x2 as y2 (for example: the label y2 is the load parameter);

[0057] Take the electrical parameters marked with categories as the training set and input them into the neural network VGG16;

[0058] Step 2.3: The training set (such as parameter x1) is input into the 3*3*C convolutional layer Conv5 of the neural network VGG16, where 3*3 represents the size of the convolutional window and C represents the number of channels. Through the convolution operation, feature extraction of electrical parameters is performed to obtain the feature map x2 output by the convolutional layer;

[0059] Step 2.4: The feature map x2 is input into the Bi-LSTM (Bi-directional Long Short-Term Memory) layer. Each row of the feature map x2 is cyclically connected to the Bi-LSTM layer to obtain a 256-dimensional Bi-LSTM, and finally the output feature map x3 is obtained;

[0060] Step 2.5: The feature map x3 is input into an encoder composed of 5 convolutional layers for encoding operations to obtain the encoded output feature x4. Successive multiplication and addition operations are performed on the encoded output feature, and then sampling is performed through the normal distribution N(0,1) to obtain the feature x6. The feature x6 is decoded by the decoder to obtain the feature x7. The feature x7 then passes through the FC fully connected layer to obtain the room coordinate position information, category score, and side compensation respectively. Based on the side compensation, the side improvement offset of the room position coordinates is obtained to improve the positioning accuracy, and the predicted electrical parameter category Y is obtained through the category score;

[0061] Step 2.6: The predicted electrical parameter category Y is compared with the previously marked correct parameter category y1, and loss calculation is performed based on the loss function to train the neural network VGG16;

[0062] Step 2.7: The electrical system diagram to be predicted is input into the trained neural network VGG16 to obtain the predicted electrical parameter categories and coordinate position information.

[0063] On the other hand, the generated knowledge graph is used for graph parameter recognition. Through the knowledge graph where entities and relationships between entities are based on design standards, knowledge graph weights are generated, and recursive neural network deep learning is introduced to improve the accuracy of electrical system diagram parameter recognition. For example, when identifying the parameters of a certain electrical device, due to considering other electrical entities connected to it (the entities in the knowledge graph contain entity information) and entity connection relationships, the knowledge graph weights are incorporated when determining the identified parameters, avoiding the misidentification of this electrical parameter as other similar electrical parameters.

[0064] Step S3: Based on the graph neural network, generate the plane layout diagram of the substation equipment;

[0065] Based on the substation room information and primary trunk line electrical parameters identified by recursive neural network deep learning, calculate the equipment grouping sequence information and generate the equipment plane layout diagram (meeting various geometric / engineering constraints), including two stages:

[0066] In the first stage, by analyzing historical power drawing elements and models, a model platform that supports national requirements, rules, standards, and models is established. At the same time, it supports the CIM (Common Information Model) specification of the IEC61970 standard, facilitating compatibility and interoperability between different power application systems.

[0067] In the second stage, using the reasoning ability of the GAT (Graph Attention Network) graph neural network, information on different relationships of entities is learned, that is, a relationship vector is added to analyze the correlation degree between different entities. Combining with the knowledge graph of the distribution network electrical field, the position, quantity, and related electrical parameters of electrical equipment in the design drawing are matched with national and industry standards, and a method for generating the plane layout diagram of substation equipment is constructed to automatically generate the plane layout diagram. The knowledge graph of the correlation degree between entities specifically includes the following steps:

[0068] Step S3.1: Map entities to the relationship space. When two entities have similar meanings, they are close in the entity space; when their meanings are not similar, they are far apart. Define the mapping operator Map and use two methods, Word2vec and the Skip-gram model, to project from the entity space to the embedded relationship space, aiming to make the distinguishability of relationships stronger.

[0069] Specifically, each relationship triple is concatenated by a connection method and then multiplied by the linear transformation parameter matrix W 1 , and converted to the relationship space C, where W 1 is learned and obtained by the Skip-gram model method, t represents the relationship type, i and j represent entity types, and i≠j;

[0070]

[0071] An entity may have multiple levels of information, and different relationships may focus on different levels of the entity. The optimal dimensions for expressing entities and relationships are not necessarily the same. Therefore, mapping to the same space may limit the model effect. For example: Entities with the same relationship may have multiple patterns. For example: The location containment relationship, (S 1 = relay, O 1 = substation, R 1 = located in), (S 2 = relay, O 2 = Room 102, R 1 = located in). Therefore, entity pairs with the same relationship are clustered.

[0072] Step S3.2: Add the parameter b of the activation function with a non-linear transformation ijt, similar to the GAT (Graph Attention Network) operation, calculate the weight coefficients of the correlation model using adjacent entity nodes (triples) on the graph:

[0073] a ijt = softmax jt (b ijt )

[0074] After using GAT to add edge relationships, perform multiple iterations on the entities to obtain the information and relationships between the entity nodes and their n - order neighbor entity nodes, so that all entities at different distances on the knowledge graph can more efficiently and globally share the information they carry through the edges, assisting in completing prediction and reasoning tasks. For example, for a large number of geometric constraints (non - overlapping, adjacent arrangement, parallel arrangement, etc.) and engineering constraints (minimum distance from the wall, minimum distance between equipment, minimum width of pedestrian passage, etc.) in the plane layout planning, according to the design standards of the substation, the constructed knowledge graph can optimize the rule base and achieve intelligent matching of the knowledge graph and the electrical constraint optimization rule base, thereby improving the accuracy of the generated substation equipment plane layout diagram and subsequent 3D diagrams.

[0075] Step S4: Based on the generative adversarial network, map the plane layout diagram to a 3D model diagram to generate a 3D model diagram of substation equipment;

[0076] Traditional 3D model modeling is based on the synthesis of grids and skeletons. Many of these traditional methods synthesize new objects by borrowing existing models in the CAD model library. Since simple interpolation calculations cannot solve the problems of the authenticity and diversity of synthesized 3D models, the present invention uses a generative adversarial network (GAN) and combines an auto - encoder - decoder network to achieve the mapping from 2D images to 3D models, for the purpose of generating 3D models of substation equipment. A 3D digital model including general - layout roads, building structures of distribution device buildings, plumbing systems, and electrical equipment, etc. is established using the 3D model generation method of substation equipment.

[0077] The present invention adopts a VAE (Variational AutoEncoder) generative adversarial network, including a generator G and a discriminator D. The 2D floor plan x is used as the input, and the latent vector z is obtained through the variational auto - encoder VAE and sent to the generator G. The generator G randomly samples a 300 - dimensional latent vector z from a probabilistic latent space and maps it into a 64*64*64 cube, and represents a 3D volume space with G(z). Using binary cross - entropy as the classification loss, the discriminator D outputs a confidence D(x) to determine whether the input floor plan x is real. The loss function of the model is jointly determined by the reconstruction error, the cross - entropy of the adversarial network, and the KL distance.

[0078] Image generation of 3D models of substation equipment based on VAE generative adversarial networks mainly completes 3D drawing of general layout plans, building plans, elevations, sections, plumbing and piping systems, and electrical equipment.

[0079] Based on the generated 3D model drawings, collision detection is carried out. Electrical equipment with intersections in the 3D drawings is screened. At the same time, through the identified electrical system diagram parameters and the knowledge graph constructed based on substation design standards, entity information and entity relationships of room information and electrical parameters are obtained, and the installation and deployment requirements in the substation design process are integrated to supplement the collision screening of electrical equipment. Since the knowledge graph is constructed based on substation design standards, and the substation design standards include requirements for avoiding collisions during installation and deployment of equipment of different sizes for different room structures, it is necessary to combine the knowledge graph and electrical system diagram parameters to supplement the screening of collision detection. Then, the collision detection results are located and given, and equipment openings are reserved at the collision positions. The 3D model of the substation building structure is completed through a 3D platform, mainly including buildings, upper steel frame structures, lower pile foundations, and main transformer equipment foundations, etc., and attribute values are assigned to building walls, doors, windows, etc. in strict accordance with the modeling specifications.

[0080] Through collision detection, collision problems in the project can be screened out. Design modifications and optimizations are made according to the collision inspection results to reduce collision problems between buildings, structures, pipelines, and electrical equipment. Collision detection on the 3D modeling platform can eliminate collision problems such as pipeline conflicts before construction, minimizing design errors that may be encountered during the construction stage and reducing construction delays and capital waste caused by collision problems.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying a substation neural network based on an electrical industry knowledge graph, characterized in that it includes the following steps: Step S1: Based on the substation design standards, construct a knowledge graph of substation design projects; Step S2: Based on the deep learning of recurrent neural networks, intelligently identify the parameters of the electrical system diagram; the parameters of the electrical system diagram include room information and primary trunk line electrical parameters. The process of intelligent identification includes the following steps: Step S2.1: Obtain the electrical system diagram and perform image segmentation. Each small segmented image contains one electrical parameter; Step S2.2: Mark the electrical parameter x1 in the small image, mark the correct category, and use the marked electrical parameter as the training set and input it into the neural network; Step 2.3: The neural network performs feature extraction of the electrical parameters through convolution operations to obtain the feature map x2 output by the convolutional layer; Step 2.4: The feature map x2 is input into the bidirectional long short-term memory layer. Each row of the feature map x2 is cyclically connected to the bidirectional long short-term memory layer, and finally the output feature map x3 is obtained; Step 2.5: The feature map x3 performs an encoding operation to obtain the encoded output feature x4. Successively perform multiplication and addition operations on the encoded output feature, and then perform normal distribution sampling to obtain the feature x6. The feature x6 is decoded to obtain the feature x7, and then through the fully connected layer, the room coordinate position information, category score, and side compensation are obtained respectively. Based on the side compensation, the side improvement offset of the room coordinates is obtained, and the predicted electrical parameter category Y is obtained through the category score; Step 2.6: Compare the predicted electrical parameter category Y with the marked correct category, calculate the loss based on the loss function, and train the neural network; Step 2.7: Input the electrical system diagram to be predicted into the trained neural network to obtain the predicted electrical parameter categories and room coordinate position information; Step S3: Use the parameters of the electrical system diagram to learn the electronic parameters as entities and the relationships between them through a graph neural network. Combine with the knowledge graph of substation design projects, match the positions, quantities, and relevant electrical parameters of electrical equipment with the standards, and automatically generate a floor plan based on the matching information; Step S4: Based on the generative adversarial network, map the floor plan to a 3D model diagram, and complete the modeling of the substation building structure model through a 3D platform to generate a 3D model diagram of substation equipment.

2. The method for identifying a substation neural network based on an electrical industry knowledge graph according to claim 1, characterized in that: In the step S3, the combination of the electronic parameters as entities and the relationships between them and the knowledge graph of substation design projects includes the following steps: Step S3.1: Map the entity space to the relationship space. When two entities have similar meanings, the distance is close; when they do not have similar meanings, the distance is far; Step S3.2: Add the parameters of the activation function of the non-linear transformation, and calculate the correlation model weight coefficient with the adjacent entity nodes on the knowledge graph: a ijt = softmax jt (b ijt ) softmax jk It represents the activation function of the j-th entity based on the relationship t, where t represents the relationship type, and i and j represent entity types. After adding the relationship of the edge using the graph neural network, the entities are iterated multiple times to obtain the information and relationships between the learned entity nodes and their n-th order neighbor entity nodes.

3. The method for identifying a substation neural network based on an electrical industry knowledge graph according to claim 2, characterized in that: In the step S3.1, each relational triple is concatenated by a connection method and then multiplied by a linear transformation parameter matrix W 1 , and is transformed into a relational space: Among which W 1 is learned and obtained by the Skip-gram method of the skip-gram model, t represents the relationship type, i and j represent entity types, and i≠j; entity pairs with the same relationship are clustered.

4. The method for identifying a transformer substation neural network based on an electrical industry knowledge graph according to claim 1, characterized in that: A variational autoencoder generative adversarial network is adopted, including a generator G and a discriminator D. The two-dimensional plane layout diagram x is used as the input. The latent vector z is obtained through the variational autoencoder and sent into the generator G. The generator G samples a multi-dimensional latent vector z from a probabilistic latent space, maps it into a cube, and represents it with a three-dimensional volume space G(z). The discriminator D outputs a confidence D(x) based on the input of the generator G to determine whether it matches the input two-dimensional plane layout diagram x. The adversarial network is trained using binary cross-entropy as the classification loss.

5. The method for identifying a transformer substation neural network based on an electrical industry knowledge graph according to claim 1, characterized in that: Based on the generated three-dimensional model diagram, collision detection is carried out, and the electrical equipment with intersections in the three-dimensional diagram is screened. At the same time, through the identified electrical system diagram parameters and the knowledge graph constructed based on the transformer substation design standards, the entity information and entity relationships of room information and electrical parameters are obtained, and the installation and deployment requirements in the transformer substation design process are integrated to supplement and screen the collisions of electrical equipment, and the collision situations are adjusted.

6. The method for identifying a transformer substation neural network based on an electrical industry knowledge graph according to claim 1, characterized in that: In the step S1, according to the relevant design standards of the transformer substation, a rule library is established, the entities in the rule library and the relationships between the entities based on the design standards are obtained, and a knowledge graph is constructed; The method of knowledge fusion is adopted. From the perspective of similarity, it is judged whether two entities are the same entity or whether they can be merged into the classification of the same entity object; the judgment based on similarity is to judge the similarity of the character descriptions, attributes / attribute values, and semantic structures of the two entities. The greater the similarity, the greater the possibility that the two entities are the same entity, or the more likely these two entities point to the same entity object.

7. The method for identifying a transformer substation neural network based on an electrical industry knowledge graph according to claim 1, characterized in that: In the step S1, the knowledge graph is represented by a triple K(S, O, R), where S represents the subject, O represents the object, and the relationship is defined as R:={SR|Value}. The relationship has a value attribute and is irreversible, that is, the subject S and the object O cannot be reversed.

Images