Distribution network middle station cross-domain data fusion method and system based on knowledge graph

Abstract

The invention discloses a cross-domain data fusion method and system for distribution network middle stations based on a knowledge graph, and particularly relates to the technical field of data processing and fusion based on the knowledge graph. Firstly, intelligent gateway equipment is adopted to collect operation, fault, maintenance and other data of all systems and all lines related to distribution network operation, the data are transmitted to a data processing and storage module through a 5G communication protocol port transmission technology, and data cleaning, data integration, data conversion and other operations are carried out; classifying and storing the data, and then constructing a mapping knowledge domain in the distribution network; according to the method, a 5G communication protocol port transmission technology and a composite model are specifically adopted to perform knowledge extraction and knowledge fusion of graph construction, so that the technical problems of data islands and other data processing in a distribution network are solved, data sharing, service sharing and resource integration are effectively realized, and the reliability of the distribution network is improved. The problem that data between distributed systems in the current distribution network field is difficult to cooperate is solved, semantic integration and sharing of data between multiple systems are achieved, and the operation and maintenance efficiency and the intelligent level of a power distribution network are improved.

Description

Technical Field

[0001] The present invention relates to the field of data processing and fusion technology based on knowledge graphs, and in particular to a cross-domain data fusion method and system for a distribution network middle station based on knowledge graphs. Background Art

[0002] In modern power systems, distribution networks are a vital link. Distribution networks receive electricity from transmission networks through distribution facilities and distribute it to users. These distribution facilities include overhead lines, cables, poles, transformers, switches and compensator equipment. In traditional distribution network management, there are challenges such as the diversity of power equipment, data heterogeneity and system complexity. There are serious data islands between subsystems, and the data processing system lacks integration and intelligence. In addition, there are limitations on data privacy, security and system architecture, which makes cross-domain data integration and processing difficult to carry out efficiently and cannot meet the needs of different data sources, departments and application scenarios.

[0003] Each subsystem of the distribution network operates independently, and each collects and processes data, resulting in data being scattered in different systems, making it difficult to share and comprehensively utilize data, and causing data island problems; secondly, the data formats generated by different systems are different, lacking standardization, which increases the difficulty of data integration. For example, the data formats of equipment operation data, fault data, maintenance records, etc. are different, which brings huge challenges to unified management; due to the complex system architecture in the distribution network, involving multiple fields and departments, the interoperability between systems is poor, making it difficult to achieve effective collaborative work; in addition, the original data of the cross-domain data in the distribution network may have noise, redundancy, and incompleteness, which affects the accuracy and effectiveness of data analysis, and how to ensure the security and privacy of data during data transmission and storage becomes an important issue, especially when it comes to sensitive user information and key equipment data; finally, traditional data processing systems are often inefficient when dealing with massive data, lack intelligent means, and are difficult to provide valuable decision support in a timely manner. Summary of the invention

[0004] In order to overcome the above-mentioned technical problems in the processing and integration of distribution network data, the present invention provides a cross-domain data fusion method and system for a distribution network middle station based on a knowledge graph, which solves the technical problems of data islands and other data processing, effectively realizes data sharing, service sharing and resource integration, and solves the problem of data difficulty in coordination between distributed systems in the current distribution network field, realizes semantic integration and sharing of data between multiple systems, and improves the operation and maintenance efficiency and intelligence level of the distribution network.

[0005] To achieve the above objectives, the present invention applies cross-domain data integration technology, big data processing and analysis technology, and intelligent operation and maintenance technology, etc., by constructing a cross-domain data knowledge graph for the distribution network middle station. The technical solution is as follows:

[0006] A cross-domain data fusion method for distribution network platforms based on knowledge graph specifically includes the following steps:

[0007] S1. First, use the intelligent gateway device to collect data from various devices and lines involved in the distribution network operation, and aggregate them into the data collection module of the distribution network center;

[0008] S2. The data in the data acquisition module of the distribution network middle station is transmitted to the data processing and storage module through the 5G communication protocol port, and operations such as data cleaning, data integration and data conversion are performed, and the data is classified into structured data, semi-structured data and unstructured data. A dedicated storage device or database management system is used to store the structured data, and then the structured data is transmitted to the knowledge graph construction module together with the semi-structured data and unstructured data;

[0009] S3, after receiving the data, the knowledge graph construction module extracts knowledge from the semi-structured data and unstructured data, and then performs knowledge fusion processing with the structured data;

[0010] S4. After being processed by the knowledge graph construction module, the data of the distribution network middle station is stored in the database, and the constructed cross-domain data knowledge graph is visualized through the terminal display module to more comprehensively reflect the cross-domain data relationship of the distribution network middle station, understand the operating status of the power system, and achieve more intelligent and reliable distribution network management.

[0011] Furthermore, the intelligent gateway device integrates multiple Internet communication methods to collect real-time data on various devices, lines, etc. involved in the distribution network operation. The device usually includes a communication module that supports multiple communication protocols, a processor module that processes data and instructions sent by IoT devices, a storage module that stores processed and analyzed data and communication data between devices, and a security module that provides data security protection and safeguards.

[0012] Furthermore, the 5G communication protocol port adopts the SA mode, and applies the interface between NG-RAN and 5G core network (5GC) to the cross-domain data fusion system of the distribution network center based on the knowledge graph, separates the system control end and the distribution network end, and conducts data communication connection through the protocol stack interface of both parties.

[0013] Furthermore, the dedicated storage device or database uses an appropriate database management system (such as MySQL, MongoDB) for storage to ensure efficient storage and retrieval, and during the storage process, data partitioning, indexing and other methods can be considered to improve retrieval efficiency.

[0014] Furthermore, in a preferred embodiment, the knowledge extraction is to extract entities, relations and attributes from semi-structured and unstructured data, and the knowledge extraction is performed using a Bert+PGD+BiLSTM+CRF combined model, as follows:

[0015] BERT (Bidirectional Encoder Representation from Transformer) is pre-trained with the Projected Gradient Descent (PGD) algorithm. The vector is input into the Bi-directional Long Short-Term Memory (Bi LSTM) model. After being encoded by the hidden layer of the Bi LSTM network, the emission matrix is ​​input into the Conditional Random Field (CRF) layer. The CRF layer serves as the decoding layer of the Bi LSTM model. The rationality of the predicted label sequence is considered in the conditional random field model, and the prediction results of the BiLSTM model can be corrected.

[0016] In a preferred embodiment, the knowledge fusion refers to integrating knowledge from different sources, resolving knowledge conflicts and duplications, and improving the consistency and integrity of knowledge. It mainly includes entity recognition and linking, knowledge merging (repeated entity merging, relationship fusion), and using a graph convolutional neural network (GCN) to operate on the entities of the knowledge graph. Its input is the feature vector and graph structure of the node (entity). Its goal is to learn the feature function of the input graph and generate node-level output. It is worth mentioning that GCN can perform end-to-end learning on input graphs of any size and shape. After performing convolution operations on the entities of the knowledge graph, GCN performs hierarchical sampling on the k-order neighbors of any central entity to form a subgraph, applies the self-attention mechanism to the subgraph, calculates the Attention value of the entity in the subgraph through the Query matrix, Key matrix and Value matrix, and finally obtains the final central entity representation through weighted average.

[0017] Assume that there are N entities in the alarm knowledge graph, and use one-hot encoding to initialize the feature representation vector of each entity, with a dimension of d. Define the matrix H = N × d, and call H the feature vector matrix. Each entity forms an N × N matrix, and define the matrix A = N × N, and call A the adjacency matrix. The GCN is shown in the following formula:

[0018]

[0019] Where: is the degree matrix of the adjacency matrix; W (l) is the learnable weight matrix; l is the number of layers of the graph convolutional network; σ is the activation function, H (l) It is the output of the previous layer of GCN and is regarded as the feature value of each node.

[0020] Furthermore, the knowledge stored in the database refers to a graph database that stores data in the form of a graph: a graph consists of nodes (entities) and edges (relationships), and the data is organized and connected through the relationship between nodes and edges. Nodes usually represent entities (such as circuit breakers, disconnectors, poles, etc.), while edges represent the relationship between these entities (such as relationships, connections, interactions, etc.). Each node and edge can have one or more properties, such as time of commissioning, manufacturer, etc. In addition, nodes can be grouped by labels, which provides greater convenience for information classification.

[0021] A distribution network middle station cross-domain data fusion system based on knowledge graph applied to the above method includes: a distribution network middle station data acquisition module, which is used to collect data from various departments, systems, and lines related to distribution network operation, and transmit it to a data processing and storage module;

[0022] The data processing and storage module is used to receive the distribution network middle station data transmitted by the distribution network middle station data acquisition module, and perform pre-processing such as cleaning and classification on the data, and then store the pre-processed data in a differentiated manner and transmit it to the knowledge graph construction module;

[0023] The knowledge graph construction module is used to receive the processed cross-domain data of the distribution network center, and use multiple models to extract entity relationships and fuse knowledge to build the knowledge graph of the distribution network center;

[0024] The storage and terminal display module is used to store the entities and relationships of the distribution network center knowledge graph and display them visually.

[0025] Furthermore, storing data in the data processing and storage module means storing structured data in a MySQL database first, and transferring semi-structured and unstructured data to the next module for processing.

[0026] Furthermore, in the knowledge graph construction module, Bert+PGD+BiLSTM+CRF is used to extract entity relationships, and knowledge fusion is achieved through GCN combined with Self-Attention to eliminate the ambiguity between entity referents and entity objects, and then store and display them.

[0027] Furthermore, the knowledge storage uses the graph database NebulaGraph to store data, which can carry ultra-large-scale data sets of hundreds of billions of points and trillions of edges, and provide millisecond-level queries. It consists of three services: Meta service, Graph service and Storage service, which is an architecture that separates storage and computing.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] The cross-domain data fusion method and system of the distribution network middle station based on knowledge graph of the present invention collects distribution network data, cleans, classifies and stores it through the data processing module, and then transmits it to the knowledge graph construction module for entity relationship extraction and knowledge fusion, and finally stores and visualizes the data through NebulaGraph. The system effectively breaks the data silos, improves data integration and sharing capabilities, improves the intelligence and efficiency of data processing, ensures data security and privacy protection, and promotes the intelligent management and decision support of the distribution network.

[0030] The present invention can be widely used in power equipment monitoring, fault diagnosis, equipment maintenance, energy efficiency management and other aspects. Through the efficient integration and intelligent analysis of massive data, it is possible to fully grasp the operating status of the distribution network and improve the operating efficiency and reliability of the distribution network. At the same time, the system can also provide important technical support for the intelligent upgrade of the distribution network and promote the digital transformation and intelligent development of the power industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0032] Figure 1 This is a flow chart of a cross-domain data fusion method for distribution network platforms based on knowledge graphs of the present invention.

[0033] Figure 2 This is a block diagram of a cross-domain data fusion system for distribution network platforms based on knowledge graphs of the present invention.

[0034] Figure 3 This is a flow chart of the implementation process and method of the cross-domain data fusion system of the distribution network center based on the knowledge graph of the present invention.

[0035] Figure 4 This is a model structure diagram for the joint extraction of entity relationships in the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0037] Example

[0038] like Figure 1 As shown, a cross-domain data fusion method for distribution network middle station based on knowledge graph specifically includes the following steps:

[0039] S1. First, use the intelligent gateway device to collect data from various devices and lines involved in the distribution network operation, and aggregate them into the data collection module of the distribution network center;

[0040] S2. The data in the data acquisition module of the distribution network middle station is transmitted to the data processing and storage module through the 5G communication protocol port, and operations such as data cleaning, data integration and data conversion are performed, and the data is classified into structured data, semi-structured data and unstructured data. A dedicated storage device or database management system is used to store the structured data, and then the structured data is transmitted to the knowledge graph construction module together with the semi-structured data and unstructured data;

[0041] S3, after receiving the data, the knowledge graph construction module extracts knowledge from the semi-structured data and unstructured data, and then performs knowledge fusion processing with the structured data;

[0042] S4. After being processed by the knowledge graph construction module, the data of the distribution network middle station is stored in the database, and the constructed cross-domain data knowledge graph is visualized through the terminal display module to more comprehensively reflect the cross-domain data relationship of the distribution network middle station, understand the operating status of the power system, and achieve more intelligent and reliable distribution network management.

[0043] In this embodiment, the distribution network middle station data refers to the real-time collection of the operation data, fault data, maintenance data and other data of each device and each line involved in the distribution network operation through an intelligent gateway device that integrates multiple Internet communication technologies, and then the collected data is integrated and uploaded. The device usually includes a communication module that supports multiple communication protocols, a processor module that processes data and instructions sent by IoT devices, a storage module that stores data after processing and analysis and communication data between devices, and a security module that provides data security protection and protection measures. Among them, the communication module is one of the most basic functional modules of the intelligent gateway. It is responsible for communicating with IoT devices and can support multiple communication protocols, such as Ethernet, Wi-Fi, Bluetooth, Zigbee, etc., and then wirelessly connect with various types of IoT devices to realize data collection, transmission and control; the processor module is the core component of the intelligent gateway, usually using a high-performance embedded processor (such as: i.MX8M / 8M Mini, Snapdragon 820 / 845 and Jetson TX2 / Nano / Xavier, etc.), with strong computing power and data processing capabilities, so that the intelligent gateway can pre-process and analyze device data and realize some simple local control and decision-making; the storage module can be used to store the data processed and analyzed by the intelligent gateway, as well as the communication data between devices, usually including memory, hard disk and flash memory, etc., which can realize fast storage and reading of data, so that the intelligent gateway can realize functions such as data caching, offline processing and long-term storage; the security module usually includes functions such as identity authentication, access control and vulnerability repair, which can ensure that the communication and data transmission between IoT devices are safe and reliable.

[0044] In this embodiment, the 5G communication protocol port adopts the SA mode, and the interface between NG-RAN and 5G core network (5GC) is applied to the cross-domain data fusion system of the distribution network center based on the knowledge graph, and the system control end and the distribution network end are separated, and data communication connection is performed through the protocol stack interface of both parties. Among them, the SA mode is a 5G networking mode, with 5G base stations, high bandwidth, low latency, and large connections, which can give full play to the advantages of 5G network; NG-RAN is a wireless network accessed by 5G, and its corresponding network element is gNB; 5G core network (5GC) manages and controls the network communication part of the system and transmits the received data.

[0045] In this embodiment, the dedicated storage device or database uses the database management system MySQL for storage to ensure efficient storage and retrieval, and in the storage process, data partitioning, indexing and other technologies may be considered to improve retrieval efficiency. Among them, MySQL is one of the most popular relational database management systems. The relational database stores data in different tables instead of putting all data in a large warehouse, which increases speed and flexibility. The SQL language used by MySQL is the most commonly used standardized language for accessing databases.

[0046] In this embodiment, the knowledge extraction is to extract entities, relations and attributes from semi-structured and unstructured data, and the Bert+PGD+BiLSTM+CRF combined model is used for knowledge extraction, as follows:

[0047] BERT (Bidirectional Encoder Representation from Transformer) is pre-trained with Projected Gradient Descent (PGD) algorithm, and the vector is input into the Bi-directional Long Short-Term Memory (Bi LSTM) model. After being encoded by the hidden layer of the Bi LSTM network, the emission matrix is ​​input into the Conditional Random Field (CRF) layer. The CRF layer is used as the decoding layer of the Bi LSTM model. The rationality of the predicted label sequence is considered in the conditional random field model, and the prediction results of the BiLSTM model can be corrected.

[0048] BERT layer: The BERT model is mainly composed of a bidirectional Transformer encoder structure. It uses the Mask pre-training task to learn the contextual semantic information of the token, and uses the next sentence prediction task to learn the sentence order feature information. BERT performs word segmentation on the input information and uses the Word Piece word segmenter to divide the original text into word blocks. We add the [CLS] mark at the beginning of the sequence to indicate the beginning of the text; add [SEP] at the end of each sentence to indicate the separation between sentences and distinguish the two input sentences before and after. By looking up the word list, each word block is mapped to a fixed word vector. If a word block is not in the word list, a special [UNK] word vector is used. In order to integrate the position information of the sequence into the word vector, BERT also embeds a position vector at each position, so that the word block at different positions will be mapped to different word vectors. At this point, the BERT model will obtain the word vector representation of each position after multiple layers of calculation on the input sequence. The input vector mainly consists of three parts: Token Embeddings (word vector), Segment Embeddings (sentence vector), and Position Embeddings (position vector). Token Embeddings divides the input words into a limited set of common subword units, that is, pay attention to the position of the separator; Segment Embeddings is used to distinguish two sentences and give them different tags; Position Embeddings encodes the position order of each word in a sentence, and then inputs the processed vector into the BiLSTM layer. BiLSTM layer: BiLSTM is a bidirectional LSTM method, which is composed of a forward LSTM and a backward LSTM. At each moment, there are two memory units to store the information of the previous and next contexts. The word vector representation containing context information is used as input to extract the features of the input sequence, and the implicit features are calculated through a bidirectional long short-term memory gate. The results are spliced ​​to obtain the text feature information further extracted by the deep learning method as the output of the hidden layer.

[0049] Assumptions is the hidden layer state of the forward LSTM network at time t, which can be regarded as a single-layer LSTM network. Calculate time t The process of x t The input at time t is:

[0050]

[0051] Assumptions is the hidden layer state of the backward LSTM network at time t, which can be regarded as a single-layer LSTM network. Calculate time t The process of x t The input at time t is:

[0052]

[0053] The final hidden state of BiLSTM is obtained by concatenating the forward sequence generated by the forward LSTM and the reverse sequence generated by the backward LSTM:

[0054]

[0055] CRF layer: As an annotation layer, it performs sentence sequence annotation. Given a training set x and a corresponding tag sequence y, as well as multiple feature functions, the CRF model parameters λ need to be learned. j ,u k And the conditional probability distribution P(y|x), the parameterized form of the conditional probability and model parameters is as follows:

[0056]

[0057] Among them, Z is the normalization factor. It can be seen that CRF has two types of feature functions. One type is the node feature function defined on the y node. This type of feature function is only related to the current node as shown in the following formula, where K is the total number of node feature functions defined on the node, and i is the position of the current node in the sequence:

[0058] s k (y i ,x,i),k=1,2,...,K

[0059] The other type is the local feature function defined in the context of y. This type of feature function is only related to the current node and the previous node as shown in the following formula, where J is the total number of local feature functions defined at the node, and i is the position of the current node in the sequence:

[0060] t j (y i+1 ,y i ,x,i),j=1,2,...,J

[0061] Based on BiLSTM feature learning, the CRF model is used to learn constraint rules. The global optimal label sequence is obtained by considering the dependency between adjacent entity labels, and the output of the BiLSTM network layer is corrected to reduce the possibility of the model outputting illegal sequences, make up for the shortcomings of the BiLSTM layer, avoid unreasonable label predictions, and realize the accurate extraction of cross-domain data entities and relationships in the distribution network center.

[0062] CRF is a probabilistic undirected graph model that solves the problems of long-term dependency and annotation bias. Its path score formula is as follows:

[0063]

[0064] Where X is the input sequence X = (x1, x2, ..., x n ); Y is the label sequence to be output Y = (y1, y2, ..., y n ); is the label transfer probability in the probability transfer matrix; P is the output score matrix, the size is n×k, where n is the number of words and k is the number of labels. Output label y for the i-th word i After obtaining the score of sequence Y, the predicted probability corresponding to label sequence Y is expressed as:

[0065]

[0066] in, represents the true and correct label sequence, Y X represents the set of all possible tags, S(X,Y) is the score function for predicting sequence Y;

[0067] Then use the dynamic programming algorithm to find the path y with the highest score * As the optimal path for sequence labeling:

[0068] y * =argmaxS(X,Y)

[0069] After the annotation layer, the model will output the results of entity relationship recognition and extraction.

[0070] Considering that the cross-domain data of the distribution network station includes data on the operation and inspection of various types of equipment, such data contains mixed numbers, Chinese and English abbreviations, and may even contain typos, and errors are inevitable when non-professionals perform annotations. These special or bad data become noise interference during the training process, affecting the learning process of the model, causing the model to have a high confidence in wrong decisions, and ultimately having a negative impact on the recognition and practical application of traditional extraction models. In addition, the scale of entity relationship extraction data sets in the power field is limited. The model can easily recognize known words, but it is difficult to understand unfamiliar words, and there is an overfitting problem, which further reduces the accuracy of actual model extraction.

[0071] The present invention introduces adversarial training in the process of entity relationship recognition and extraction. By adding disturbances to the embedding layer word vector of the model as perturbation samples and training together with the original samples, the overall learning and generalization capabilities of the model are improved, better training effects are achieved, and higher-precision recognition and extraction are realized.

[0072] PGD ​​adversarial layer: Adversarial training introduces excessive interference noise, which can easily destroy the effective features of the data and reduce model performance. The rapid rise of gradients can easily lead to a simple and crude "one-step" phenomenon, which makes it impossible to obtain the optimal point within the constraints. Therefore, the PGD iterative method is adopted. Through the iterative strategy of "small steps, more steps", a fast search for the optimal point is achieved while ensuring a certain level of word vector disturbance:

[0073] By calculating the forward loss value and obtaining the gradient through back propagation, the embedding layer weights are adversarially perturbed according to the adversarial training formula and the initial gradient, and then the loss value and gradient after the attack are calculated according to the model weights updated with the parameters after the attack to determine whether they are within the constraint space. If not, the initial weights of the embedding layer are restored, and the original gradients and the gradients after the attack are accumulated to update the model weights. Otherwise, adversarial perturbations are performed again, and the loss value and gradient after the attack are calculated according to the model weights updated with the parameters after the attack until it is determined that they are not within the constraint space. The model weights are updated for adversarial training, and the adversarial samples are input into BiLSTM for training. Finally, the optimal sequence label is output through the CRF layer.

[0074] In this embodiment, the knowledge fusion refers to integrating knowledge from different sources, resolving knowledge conflicts and duplications, and improving the consistency and integrity of knowledge. It mainly includes entity recognition and linking, knowledge merging (repeated entity merging, relationship fusion), and adopts a graph convolutional neural network (GCN) to operate on the entities of the knowledge graph. Its input is the feature vector and graph structure of the node (entity). Its goal is to learn the feature function of the input graph and generate node-level output. GCN can perform end-to-end learning on input graphs of any size and shape.

[0075] After performing convolution operations on the entities in the knowledge graph, GCN hierarchically samples the k-order neighbors of any central entity to form a subgraph, applies the self-attention mechanism to the subgraph, calculates the Attention value of the entity in the subgraph through the Query matrix, Key matrix and Value matrix, and finally obtains the final central entity representation through weighted average.

[0076] Assume that there are N entities in the alarm knowledge graph, and use one-hot encoding to initialize the feature representation vector of each entity, with a dimension of d. Define the matrix H = N × d, and call H the feature vector matrix. Each entity forms an N × N matrix, and define the matrix A = N × N, and call A the adjacency matrix. The GCN is specifically shown in formula (1):

[0077]

[0078] Where: is the degree matrix of the adjacency matrix; W (l) is the learnable weight matrix; l is the number of layers of the graph convolutional network; σ is the activation function, H (l) It is the output of the previous layer of GCN and is regarded as the feature value of each node.

[0079] Perform hierarchical sampling on the k-order neighbors of the central entity e0 to form a subgraph, and form the entities in the subgraph into a sequence (e0, e1, ..., e m ), the Self-Attention mechanism is used to calculate the attention value of the subgraph central entity e0 and its k-order neighbor entities, that is, the entities in the sequence, and finally the feature vector representation of the central entity is obtained through weighted sum. The specific formula is as follows:

[0080]

[0081] α 0,i =(q 0 ) T ·k i

[0082]

[0083] In the formula, is an entity sequence (e0,e1,...,e m ) in a certain entity feature vector; W q , W k , W v They are respectively the learnable query, key, and value weight matrices, through which the query vector q, key vector k, and value vector v of the entity are obtained respectively; is the normalized attention value related to the central entity; is the final center entity vector representation obtained by weighted sum.

[0084] On this basis, the similarity between the entity representation in the input text and the entity representation in the knowledge graph is calculated, and the one with the highest similarity score is selected for linking.

[0085] In this embodiment, the knowledge stored in the database refers to a graph database that stores data in the form of a graph: a graph is composed of nodes (entities) and edges (relationships), and data is organized and connected through the relationship between nodes and edges. Nodes usually represent entities (such as circuit breakers, disconnectors, poles, etc.), while edges represent the relationship between these entities (such as relationship, connection, interaction, etc.). Each node and edge can have one or more properties, such as time of commissioning, manufacturer, etc. In addition, nodes can be grouped by labels, which provides greater convenience for information classification;

[0086] Embodiment 2:

[0087] like Figure 2 As shown, a knowledge graph-based distribution network middle station cross-domain data fusion system applied to the method of embodiment 1 includes:

[0088] The data collection module of the distribution network center is used to collect data from various departments, systems and lines related to the distribution network operation and transmit it to the data processing and storage module;

[0089] The data processing and storage module is used to receive the distribution network middle station data transmitted by the distribution network middle station data acquisition module, and perform pre-processing such as cleaning and classification on the data, and then store the pre-processed data in a differentiated manner and transmit it to the knowledge graph construction module;

[0090] The knowledge graph construction module is used to receive the processed cross-domain data of the distribution network center, and use multiple models to extract entity relationships and fuse knowledge to build the knowledge graph of the distribution network center;

[0091] The storage and terminal display module is used to store the entities and relationships of the distribution network center knowledge graph and display them visually.

[0092] Storing data in the data processing and storage module means storing structured data in the MySQL database first, and transferring semi-structured and unstructured data to the next module for processing.

[0093] In the knowledge graph construction module, Bert+PGD+BiLSTM+CRF is used to extract entity relationships, and knowledge fusion is achieved through GCN combined with Self-Attention to eliminate the ambiguity between entity referents and entity objects, and then store and display them.

[0094] The data processing classification is to classify the processed distribution network middle station data into structured data, semi-structured data, and unstructured data according to the corresponding requirements such as data format, required storage method, query and operation method. Among them, structured data has a strict table form and fixed fields, and is stored in a relational database; semi-structured data contains partial structural information such as labels or key-value pairs, with a flexible format, commonly found in JSON, XML and other files; unstructured data has no predefined structure, including text, images, audio, video, etc. In the storage and terminal display module, the graph database used is NebulaGraph, which can carry ultra-large-scale data sets of hundreds of billions of points and trillions of edges, and provide millisecond-level queries. It consists of three services: Meta service, Graph service and Storage service, which is an architecture that separates storage and computing. Specifically, the Meta service is a housekeeper, and its main function is data management, such as Schema operation, cluster management and user authority management; the Graph service is responsible for processing query requests, including parsing query statements, verifying statements, generating execution plans and executing according to execution plans; the Storage service is responsible for the storage of specific data;

[0095] NebulaGraph uses RocksDB as the local storage engine for data storage and implements its own KVStore. The main data stored is points and edges. NebulaGraph stores the information of points and edges as keys, and stores the attribute information of points and edges in values, so that attribute filtering can be used more efficiently. In order to improve query performance, NebulaGraph supports creating indexes for the tag of points or a certain attribute of tags, the edge type of edges or a certain attribute of edge types. Data can be queried based on the specified tags, edge types, and attributes. However, the index itself does not store data, but stores the location of data. It is only used to query scenarios based on pure attribute conditions. Queries that start from a point and expand outward do not require indexes based on VID queries. NebulaGraph supports two types of indexes: native indexes (stored and sharded together with point data locally, and data can be queried based on specified attributes) and full-text indexes (implemented based on Elasticsearch, used for prefix search, wildcard search, regular expression search, and fuzzy search on string attributes).

[0096] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A cross-domain data fusion method for distribution network platforms based on knowledge graph, characterized in that: The following steps are involved: S1. First, use the intelligent gateway device to collect data from each device and line involved in the distribution network operation, and aggregate them into the data collection module of the distribution network center; S2. The data in the data acquisition module of the distribution network middle station is transmitted to the data processing and storage module through the 5G communication protocol port, and operations such as data cleaning, data integration and data conversion are performed, and the data is classified into structured data, semi-structured data and unstructured data. A dedicated storage device or database management system is used to store the structured data, and then the structured data is transmitted to the knowledge graph construction module together with the semi-structured data and unstructured data; S3, after receiving the data, the knowledge graph construction module extracts knowledge from the semi-structured data and unstructured data, and then performs knowledge fusion processing with the structured data; S4. After being processed by the knowledge graph construction module, the data of the distribution network middle station is stored in the database, and the constructed cross-domain data knowledge graph is visualized through the terminal display module to more comprehensively reflect the cross-domain data relationship of the distribution network middle station, understand the operating status of the power system, and achieve more intelligent and reliable distribution network management.

2. According to claim 1, the cross-domain data fusion method for distribution network platforms based on knowledge graph is characterized by: In step S1, the intelligent gateway device integrates multiple Internet communication methods to collect real-time data of various devices and lines involved in the distribution network operation. The device includes a communication module that supports multiple communication protocols, a processor module that processes data and instructions sent by IoT devices, a storage module that stores processed and analyzed data and communication data between devices, and a security module that provides data security protection and safeguards.

3. The cross-domain data fusion method for distribution network platforms based on knowledge graph according to claim 1 is characterized in that: In step S2, the 5G communication protocol port adopts the SA mode, and the interface of NG-RAN and 5G core network 5GC is applied to the cross-domain data fusion system of the distribution network center based on the knowledge graph, so as to separate the system control end and the distribution network end, and establish data communication connection through the protocol stack interface of both parties.

4. According to claim 1, the cross-domain data fusion method for distribution network platforms based on knowledge graph is characterized by: In step S2, the dedicated storage device or database is stored using a database management system to ensure efficient storage and retrieval, and during the storage process, data partitioning and indexing methods may be considered to improve retrieval efficiency.

5. According to claim 1, the cross-domain data fusion method for distribution network platforms based on knowledge graph is characterized by: In step S3, the knowledge extraction is to extract entities, relations and attributes from semi-structured and unstructured data, and the Bert+PGD+BiLSTM+CRF combined model is used for knowledge extraction, as follows: BERT (Bidirectional Encoder Representation from Transformer) is pre-trained with the Projected Gradient Descent (PGD) algorithm. The vector is input into the Bi-directional Long Short-Term Memory (Bi LSTM) model. After being encoded by the hidden layer of the Bi LSTM network, the emission matrix is ​​input into the Conditional Random Field (CRF) layer. The CRF layer serves as the decoding layer of the Bi LSTM model. The rationality of the predicted label sequence is considered in the conditional random field model, and the prediction results of the Bi LSTM model are corrected.

6. The cross-domain data fusion method for distribution network platforms based on knowledge graph according to claim 1 is characterized by: In step S3, the knowledge fusion refers to integrating knowledge from different sources, solving knowledge conflicts and duplications, and improving the consistency and integrity of knowledge. It mainly includes entity recognition and linking, knowledge merging, and adopts graph convolutional neural network (GCN) to operate on the entities of the knowledge graph. Its input is the feature vector and graph structure of the entity node. Its goal is to learn the feature function of the input graph and generate node-level output. GCN can perform end-to-end learning on graphs of any size and shape. After performing convolution operations on the entities in the knowledge graph, GCN performs hierarchical sampling on the k-order neighbors of any central entity to form a subgraph, applies the self-attention mechanism to the subgraph, calculates the attention value of the entity in the subgraph through the query matrix, key matrix and value matrix, and finally obtains the final central entity representation through weighted average sum; Assume that there are N entities in the alarm knowledge graph, use one-hot encoding to initialize the feature representation vector of each entity, the dimension is d, define the matrix H = N × d, call H the feature vector matrix, each entity forms an N × N matrix, define the matrix A = N × N, call A the adjacency matrix, GCN is specifically as shown in the formula: Where: is the degree matrix of the adjacency matrix; W (l) is a learnable weight matrix; l is the number of layers of the graph convolutional network; σ is the activation function, H (l) It is the output of the previous layer of GCN and is regarded as the feature value of each node.

7. The cross-domain data fusion method for distribution network platforms based on knowledge graph according to claim 1 is characterized by: In step S4, the knowledge is stored in a database, which refers to a graph database that stores data in the form of a graph: a graph consists of node entities and edge relationships, and data is organized and connected through the relationships between nodes and edges. Nodes represent entities, while edges represent the relationships between these entities. Each node and edge can have one or more attributes. In addition, nodes can be grouped by labels, which provides greater convenience for information classification.

8. A cross-domain data fusion system for distribution network platforms based on knowledge graph, characterized in that: include: The data collection module of the distribution network center is used to collect data from various departments, systems and lines related to the distribution network operation and transmit it to the data processing and storage module; The data processing and storage module is used to receive the distribution network middle station data transmitted by the distribution network middle station data acquisition module, and perform pre-processing such as cleaning and classification on the data, and then store the pre-processed data in a differentiated manner and transmit it to the knowledge graph construction module; The knowledge graph construction module is used to receive the processed cross-domain data of the distribution network center, and use multiple models to extract entity relationships and fuse knowledge to build the knowledge graph of the distribution network center; The storage and terminal display module is used to store the entities and relationships of the distribution network center knowledge graph and display them visually.

9. The cross-domain data fusion system for distribution network platforms based on knowledge graph according to claim 8 is characterized in that: Storing data in the data processing and storage module means storing structured data in the MySQL database first, and transferring semi-structured and unstructured data to the next module for processing.

10. The cross-domain data fusion system for distribution network platforms based on knowledge graph according to claim 8 is characterized in that: In the knowledge graph construction module, Bert+PGD+BiLSTM+CRF is used to extract entity relationships, and knowledge fusion is achieved through GCN combined with Self-Attention to eliminate the ambiguity between entity referents and entity objects, and then store and display them.

11. The cross-domain data fusion system for distribution network platforms based on knowledge graph according to claim 8 is characterized in that: The knowledge storage uses the graph database NebulaGraph to store data, which can carry ultra-large-scale data sets of hundreds of billions of points and trillions of edges, and provide millisecond-level queries. It consists of three services: Meta service, Graph service and Storage service, and is an architecture that separates storage and computing.

Images