Electric power knowledge graph construction method based on heterogeneous data source
Through a method based on heterogeneous data source, multi-source heterogeneous data is collected and processed, and semantic matching is used by deep neural networks and typical correlation analysis, which solves the problem of low efficiency and accuracy of power knowledge graph construction, and achieves efficient power Knowledge graph construction and intelligent decision support.
Patent Information
- Application Number
- CN202411084868.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-29
AI Technical Summary
The efficiency and accuracy of the existing power knowledge graph construction are not high, and it is difficult to meet the needs of power supply guarantee for large-scale events.
Using a method based on heterogeneous data sources, semantic matching is carried out by collecting multi-source heterogeneous data, unstructured data into structured data, Chinese word segmentation, deep neural network and typical correlation analysis to construct a power knowledge graph.
It improves the efficiency and accuracy of the construction of knowledge graphs, realizes efficient classification and data fusion of power knowledge graphs, and enhances the intelligent decision-making support of the power system.
Smart Images

Figure CN119047554A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a method for constructing an electric power knowledge graph based on heterogeneous data sources. Background Art
[0002] In order to better meet the needs of power supply guarantee for major events, large-scale important power supply guarantee work usually involves the construction of a collaborative power supply guarantee platform, making full use of modern information means to establish an intelligent command system covering various professions and units, and achieving unified command, online monitoring, data sharing, rapid response and collaborative disposal throughout the process. The power supply guarantee platform for major events needs to integrate all kinds of power supply guarantee data, analyze and refine these element information, and provide intelligent power supply customer service, intelligent information aggregation, and intelligent decision-making support for power supply guarantee work. The construction of knowledge graphs is an important part of this.
[0003] For example, Chinese patent CN113434634A, published on September 24, 2021, includes: determining the ontology of the target knowledge graph, the ontology consisting of entities and relationships between entities; extracting target entities and target relationships between entities from target data based on the ontology, the target entities are at least used to represent the names of each site in the power grid, as well as the operating status and maintenance information of each site; based on the target entities and target relationships, obtaining the target knowledge graph, wherein the entities in the target knowledge graph are at least used to represent the names of each site in the power grid, as well as the operating status and maintenance information of each site, and the relationships between entities in the target knowledge graph are at least used to represent the associations between each site. The knowledge graph constructed by this method is inefficient and not very accurate. Summary of the invention
[0004] The technical problem to be solved by the present invention is that the existing power knowledge graph construction efficiency and accuracy are low. A power knowledge graph construction method based on heterogeneous data sources is proposed, which can effectively improve the construction efficiency and accuracy of the knowledge graph.
[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is: a method for constructing an electric power knowledge graph based on heterogeneous data sources, comprising the following steps: S1: Collect multi-source heterogeneous data; S2: Classify the collected heterogeneous data and convert unstructured data text into structured data text; S3: Use CRF algorithm to perform Chinese word segmentation on structured data text; S4: Use canonical correlation analysis and deep neural network to perform layer-by-layer semantic matching on the obtained word segmentation data, and use the matched data to build an electric power knowledge graph.
[0006] A method for constructing an electric power knowledge graph based on heterogeneous data sources collects multi-dimensional heterogeneous data in the form of data text, converts the unstructured data text into structured data text, and then performs Chinese word segmentation through the CRF algorithm. Finally, data fusion is performed through semantic matching, and the fused data is filled into the knowledge graph framework to obtain a complete electric power knowledge graph.
[0007] Preferably, the collected heterogeneous data is divided into power grid data, user data and social data according to the source, and divided into structured data and unstructured data according to the type. According to the professional terms in the relevant fields, semantic calculation is performed on the collected heterogeneous data, so as to classify the collected heterogeneous data. Heterogeneous data is divided into power grid data, user data and social data according to the source, and divided into structured data and unstructured data according to the type.
[0008] Preferably, the converted structured data text is in a three-level format: a header indicating the data type; an index indicating the data address; and a storage indicating the data content. The unified data format facilitates the final fusion.
[0009] As a preferred method, the CRF algorithm is used in combination with pre-labeled sample data to output the words with the highest probability of appearing in the structured data text as word segmentation. Text information is usually a continuous text sequence mixed with a large number of modal particles without practical meaning, so Chinese word segmentation is required.
[0010] As a preferred method, a deep belief network is used as the basic model of a deep neural network. The deep belief network is composed of multiple restricted Boltzmann machines. The network parameters are trained layer by layer in forward and reverse directions. Canonical correlation analysis is integrated into the deep belief network to participate in the correlation analysis of source domain data and target domain data. Canonical correlation analysis (CCA) is combined with a deep neural network. Through layer-by-layer semantic matching of multi-source heterogeneous modal data after word segmentation, a deep semantic matching model that satisfies domain private networks and domain shared networks is constructed.
[0011] Preferably, semantic computing includes the following: first determine the relevant core words and core word associated word sets in different categories, distribute and calculate the sum of similarities between heterogeneous data and core words and core word associated word sets in different categories, classify heterogeneous data with a sum of similarities greater than 0, and classify heterogeneous data into the category with the largest sum of similarities. Through semantic computing, the categories of heterogeneous data are quickly distinguished, thereby improving the efficiency of knowledge graph construction.
[0012] Preferably, the layer-by-layer semantic matching includes the following: using a stacked automatic encoder to encode the source domain and the target domain, extracting the cross-domain symbiotic data through the encoding results to obtain hidden layer features; performing typical correlation analysis on the encoded source domain and target domain models, and obtaining the maximum correlation coefficient matrix between the two data domains based on the symbiotic data features. The hidden features of different data are obtained through the above method, and then fused.
[0013] Preferably, during the back propagation process of the deep belief network model, the maximum correlation coefficient matrix is used to fine-tune the domain network parameters, thereby enhancing the correlation between the source domain and the target domain and improving the final accuracy.
[0014] As a preference, the matched data is filled into the electric power knowledge graph framework according to the association between the data to complete the construction of the electric power knowledge graph. The electric power knowledge graph framework is divided into multiple levels according to the types, and the nodes between the levels are connected through relationships. The final electric power knowledge graph is a mesh hierarchical structure, and each mesh node is interconnected.
[0015] As a preferred embodiment, the electric power knowledge graph has query and update functions. The query information instruction includes the query node name and query content. When the electric power knowledge graph receives the update function, it determines the update node and updates the information of the relevant nodes according to the received information text. When the data of a certain node changes, the relevant node information in the electric power knowledge graph also needs to be updated.
[0016] The substantial effect of the present invention is as follows: the present invention designs a method for constructing an electric power knowledge graph based on heterogeneous data sources, improves the classification efficiency of heterogeneous data and the efficiency of knowledge graph construction through semantic computing methods; and fuses heterogeneous data through typical correlation analysis and deep neural networks to improve the accuracy and construction efficiency of the final knowledge graph. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a flow chart of Example 1.
[0018] Figure 2 This is a framework diagram of the overall construction method of the power knowledge graph of Example 1.
[0019] Figure 3 This is the heterogeneous transfer learning model framework of Example 1.
[0020] Figure 4 It is a comparison chart of the experimental group and the control group of Example 1.
[0021] Figure 5 This is a diagram showing the impact of the number of hidden layers and the number of neurons on the semantic error in Example 1. DETAILED DESCRIPTION
[0022] The specific implementation of the present invention will be further described below through specific examples in conjunction with the accompanying drawings.
[0023] Embodiment 1: A method for constructing a power knowledge graph based on heterogeneous data sources, such as Figure 1 and Figure 2 As shown, first of all, it is necessary to collect various heterogeneous data in the power grid. These heterogeneous data mainly include three categories: basic entity information, information association and the attributes. Basic entity information includes: electricity customers, power grid companies, power plants, electricity consumption, place names, units of measurement, and dates. Information associations include: electricity supply and demand relationship, electricity consumption size relationship, and electricity user affiliation relationship. The attributes include: geographical location, creation date, and location. In this embodiment, this information is divided into three categories: power grid data, user data, and social data. Power grid data covers the generation, transmission, fault troubleshooting and quality assessment of electric energy, involving power production systems, automatic power supply voltage acquisition systems, fault repair management systems, and data acquisition and control systems. User data mainly refers to acquisition system data, charging pile data, CMS system data, etc. Social data mainly refers to public service system data, meteorological system data, and geographic system data.
[0024] After completing the collection of heterogeneous data in the power grid, the collected heterogeneous data needs to be classified. The collected heterogeneous data is first classified into three categories: power grid data, user data, and social data, and then all heterogeneous data are divided into structured data and unstructured data. If manual classification is used to classify heterogeneous data, the workload is large and the efficiency is low. Therefore, this embodiment uses a semantic calculation method to divide it. Use common terms in various classifications as keywords to construct a word set, and construct the associated words that can be obtained by association with the keyword word set into a first-level associative word set. The associated words that can be obtained by association with the first-level associative word set are constructed into a second-level associative word set, and so on, to obtain a complete data set. Then, the similarity with the data set is calculated by the semantic association calculation formula. The calculation formula is as follows: Among them: SET(K) is the sum of similarities, and the similarity with the keyword is greater than 0; K i is the ith keyword; W j is the jth word in the i-th layer; the largest sum of similarities calculated from data sets under different categories indicates which category the heterogeneous data belongs to.
[0025] After the heterogeneous data is divided, the unstructured data text needs to be converted into structured data text. Before the conversion, the data text needs to be preprocessed. Whether it is power grid data, user data or social data, there is a large amount of text information. Text information is different from digital information. It is unstructured information and cannot be directly extracted for features; and because text information is usually a continuous text sequence and is mixed with a large number of meaningless modal particles, Chinese word segmentation is required. This embodiment adopts the CRF algorithm, uses pre-labeled sample data for model training, and then outputs the word with the highest probability as the segmentation result. The calculation process of the CRF algorithm is as follows: Where: x, y are a field and sample data in the data text to be processed respectively; Z(X) represents the normalization factor; W i 、F i is the weight coefficient; w is the final probability, which is obtained through the iteration of the gradient descent algorithm, and the field with the largest final probability is output as the word segmentation.
[0026] After completing the Chinese word segmentation of the heterogeneous data text, it is necessary to convert the unstructured data text into structured data text. In order to reduce the difficulty of converting unstructured data such as text into structured data, the multi-source heterogeneous data information model of the power system should have a unified format when performing data collection and output. In this embodiment, the data physical structure of the multi-source heterogeneous data information model of the power system is designed into a three-level form: header, index and storage.
[0027] Since the number of data instances of some modes in the power system is small, the feature extraction is inaccurate. This embodiment uses transfer learning to solve the feature learning of domains with fewer data instances. The specific process is: combining canonical correlation analysis (CCA) with deep neural networks, and constructing a deep semantic matching model that meets the domain private network and domain shared network through layer-by-layer semantic matching of multi-source heterogeneous modal data after word segmentation.
[0028] like Figure 3 As shown, a deep neural network usually includes: an input layer, a hidden layer, and an output layer, and these layers also include multiple branches. This embodiment uses a deep belief network as the basic model of the multi-source heterogeneous data fusion algorithm, which is composed of multiple restricted Boltzmann machines, and uses layer-by-layer forward and reverse training of network parameters. In this embodiment, typical correlation analysis is integrated into the deep belief network to participate in the correlation analysis of source domain data and target domain data. The deep belief network can be used as a pre-training of initial parameters to improve the performance and convergence speed of the model.
[0029] Canonical correlation analysis can map related features in different data domains to a feature subspace through a matrix, thereby extracting features with strong correlation. Considering that there are multiple data sources for power industry data and external data, and the volume of different data sources varies, canonical correlation analysis is used to extract shared features between different data sources, thereby realizing transfer learning. In this process, the symbiotic data C of different data sources S , C T The hidden layer features of are extracted under the action of the transformation matrix, and the relevant matching coefficient matrix is obtained. The specific process is as follows: The source domain and the target domain are encoded using a stacked autoencoder, and the cross-domain co-occurrence data is extracted through the encoding results to obtain the corresponding hidden layer features.
[0030] Q S =f(U S C S +b S ) Q T =f(U T C T +b T ) Among them: U S , U T Represent the network weight matrices of the source domain and the target domain respectively; C S , C T is the symbiotic data; b S 、b T Distribution is used to represent the bias vectors corresponding to the source domain and the target domain; f() represents a nonlinear function.
[0031] 2) Perform canonical correlation analysis on the encoded source domain and target domain. According to the co-occurrence data characteristics, the maximum correlation coefficient matrix U between the two data domains can be obtained. S and U T The maximum correlation coefficient matrix can be used to map the co-occurrence data features of the source and target domains into the semantic shared subspace; 3) In order to enhance the correlation between the source domain and the target domain, the parameters of the deep belief network model need to be optimized. In the back propagation process of the deep belief network model, that is, the process from the target domain to the source domain, the matrix U S and U T Domain network parameters (U S ,b S ,U T ,b T ) to make fine adjustments. Figure 3 In the transfer learning model shown, the objective function is set to: 1) minimize the reconstruction error between the source domain and the target domain; 2) maximize the correlation of the cross-domain deep network. Therefore, the expression of the objective function is: minH=H S (U S , b S )+H T (U T , b T )+p p=a(V S , V T ) Where: H S (U S , b S ), H T (U T , b T ) represent the reconstruction errors of the source domain and the target domain after encoding and training with the deep belief network. The maximum correlation coefficient matrix V between p and the source domain and the target domain S 、V T The projection vector V of the maximum correlation coefficient matrix is used to represent the correlation matching function between the two. S 、V T , the maximization of the co-occurrence data between the source domain and the target domain can be further converted into the problem of maximizing the feature correlation.
[0032] There are six parameters in the objective function: U S 、b S , U T 、n T 、V S 、V T If these six parameters are optimized at the same time, it will bring a huge amount of calculation. This paper uses Lagrange multipliers and stochastic gradient descent to optimize (U S 、b S )、(U T 、b T ) and (V S 、V T ) is optimized. Since the objective function (V S 、V T ) only has the third part, so we can give the source domain and the target domain (U S 、b S )、(U T 、b T ), for (V S 、V T ) and perform Lagrange transformation to obtain V S and V T . And (U S 、b S )、(U T 、b T ) optimization, we need to give the correlation coefficient matrix [VS , V T ]. The optimization process of the source domain and the target domain is the same, and only the optimization of the source domain is described here. In the back propagation process of deep learning training, the gradient descent algorithm is used to optimize U S , b S Make adjustments.
[0033] The knowledge graph generated by this embodiment is also configured to receive query instructions and return reply information corresponding to the query instructions, wherein the query instructions are at least used to determine the names of each node in the power grid that supplies power to the target object, and to query the operating status and maintenance information of the nodes.
[0034] Specifically, based on the above target knowledge graph, the method and principle of returning corresponding information according to the query instruction are as follows: The above-mentioned knowledge graph is also used to realize the knowledge-driven intelligent aggregation of power supply information. It can be seen from the construction process of the knowledge graph that it itself has the function of information aggregation. Based on the knowledge-driven information aggregation, the information aggregation ability of the knowledge graph itself is used to realize the aggregation of power supply information. Aggregation methods include information aggregation based on a single entity, information aggregation based on multiple entities, and knowledge aggregation for power supply tasks. A single entity in the knowledge graph is already the result of the aggregation of power supply information from multiple heterogeneous data sources. In addition, according to the relationship between entities, it is possible to wander to other adjacent entities in the knowledge graph. For the aggregation of power supply information of multiple entities, the corresponding entity information is obtained through semantic query. Task-oriented knowledge aggregation can be divided into inheritance-based framework aggregation, context-based knowledge aggregation, predicate dependency-based rule aggregation, and entity relationship-based neighbor aggregation. Inheritance-based framework aggregation directly uses the inheritance relationship between frameworks to aggregate the ancestor and child framework knowledge of all frameworks to form a task-related framework set. Context-based knowledge aggregation improves the timeliness of aggregation based on the context, and improves the accuracy of aggregation while reducing the aggregation time. Predicate dependency-based rule aggregation aggregates all rules with dependencies based on the predicate dependency path on the rule set. The neighbor aggregation based on entity relations takes the entity and relation names contained in the power supply task data as clues, and aggregates the (entity-relationship-entity) triples adjacent to the clues from the knowledge graph composed of entity-relationship.
[0035] Then, semantic query and search based on power supply protection knowledge are realized, and effective use of power supply protection knowledge is achieved through semantic query and search. Semantic query realizes complex multi-hop query, which is one of the foundations of human-computer dialogue based on power supply protection knowledge graph. Search realizes the precise positioning of power supply protection information, and further improves the accuracy of search results on the basis of the high recall rate of traditional information retrieval technology. At the same time, knowledge query service based on subject is realized.
[0036] In some embodiments of the present application, after receiving a query instruction, a target query instruction can also be generated based on multiple query instructions, wherein the target query instruction is used to query based on multiple query instructions, wherein the target query instruction is at least used to determine the names of key nodes in each node whose number of queries is higher than a preset threshold, and to query the operating status and / or maintenance information of the key nodes.
[0037] With the help of the power supply knowledge graph that connects various power supply information sources, the information push application supporting the large-scale activity command platform can be realized, and the user's interactive question and answer requests can be received in the form of voice or natural language text. With the help of natural language understanding, intelligent question and answer, semantic search and recommendation and other technologies, the natural language human-computer intelligent interaction of power supply information can be realized, avoiding the traditional rigid relational database query and retrieval, realizing friendly human-computer interaction, and realizing the full reuse of the power supply knowledge graph. In the above process, through the model based on convolutional neural network, the question and answer intention is accurately identified, the question category is analyzed, and the question and answer elements such as entities and relationships are further accurately extracted through the pre-trained language model; through the dynamic knowledge graph technology, according to the question and answer intention and question and answer elements, the most likely questions are automatically predicted and generated, and converted into queries for the knowledge graph; the question understanding and semantic query are spatially and temporally expanded, so that the intelligent question and answer provides power supply knowledge involving spatiotemporal relationships; through the prediction of subsequent questions, question and answer scene recognition, question and answer log mining, etc., multiple rounds of intelligent question and answer for power supply are realized, and question and answer scene switching can be performed.
[0038] Furthermore, the understanding and recommendation of power supply plans based on knowledge graphs can be realized, and the intelligent early warning capability of power supply risks can be formed. Power supply risk warning is crucial to the power supply business of large-scale events. As power supply data, information and brain, the power supply knowledge graph can provide explainable intelligent early warning analysis, processing and auxiliary decision support for power supply risks. By organizing the knowledge of fault handling business experts, building corresponding fault knowledge in the power supply knowledge graph, and turning power supply fault elements into knowledge, including fault causes, related factors, handling rules, decision plans, etc., as a professional knowledge base for fault prediction and fault handling; collect various types of fault-related data, extract fault features through deep learning models, train fault prediction models, and form the ability to predict faults in advance before they occur; based on the fault prediction model, give fault reminders in advance, reasonably allocate patrol manpower, and combine manual on-site analysis to eliminate power supply fault blind spots; in the process of fault occurrence and handling, quickly gather relevant power supply information, notify relevant personnel, and provide command decisions based on fault rules, and even automatically handle power supply faults.
[0039] In this embodiment, the target data changes dynamically. After obtaining the target knowledge graph, if the target data changes, the processing module also needs to determine the changed data in the target data and update the target knowledge graph based on the changed data.
[0040] Specifically, in the knowledge extraction of the power field, common reasons that lead to changes in target data include rapid iteration of new equipment, repeated experiments, pre-maintenance, emergency repairs during the event, or the occurrence of new types of failures.
[0041] When the target data changes, in order to update the knowledge graph, it is necessary to determine the target entity keywords and target relationship keywords in the changed data; then map the target entity keywords to target entities, and map the target relationship keywords to target relationships, where the target relationship is the association relationship between target entities; finally, add the target entity and target relationship to the target knowledge graph to obtain an updated target knowledge graph, and use the updated target knowledge graph as the new target knowledge graph.
[0042] In some embodiments of the present application, the target knowledge graph can be updated based on the target knowledge graph itself when the target data changes. The specific process is as follows: determine the corpus data corresponding to the target entities and target relationships in the target knowledge graph; based on the corpus data, determine the target entity keywords and target relationship keywords of the changed part.
[0043] It is also possible to determine the target entity keywords and target relationship keywords in the changed data based on the domain dictionary in the power supply protection field. The specific process is as follows: determine the target entity keywords and target relationship keywords of the changed part based on the domain dictionary, wherein the domain dictionary is used to store the corpus data corresponding to the entities and relationships in the power supply protection field, and the corpus data stored in the domain dictionary is at least partially different from the corpus data corresponding to the target knowledge graph.
[0044] The data sources of the above-mentioned domain dictionaries include the historical data accumulated previously, such as the power supply field dictionary, power supply equipment dictionary, power supply equipment maintenance dictionary, and open source power supply field knowledge data, such as power supply encyclopedia entries.
[0045] The above-mentioned method of updating the target knowledge graph based on the target knowledge graph itself and based on the power protection domain dictionary can be carried out simultaneously, and the automatic update of the target knowledge graph can be realized by machine learning. The specific process is as follows: First, the power protection data set is automatically annotated through the power protection domain dictionary. At the same time, the data set is automatically annotated using the constructed domain knowledge graph. In this remote supervision process, the generation of noise data can be suppressed in the following way: a statistical learning method is used to perform word segmentation and statistical analysis on the annotated corpus accumulated by the project, and then a dictionary oriented to corpus information is constructed in combination with a manual verification method. The dictionary only includes words that are useful for determining knowledge; then the data obtained by automatic annotation is preliminarily verified. It only needs to segment the annotated data and then scan the words contained in the data. If the data does not contain any "useful" words, it can be preliminarily determined to be noise and abandoned.
[0046] Then, in order to further ensure the quality of the labeled data results, unreliable labeling results are filtered out through a voting mechanism, and a reinforcement learning mechanism is introduced in subsequent model training to reward correct labeling, punish incorrect labeling, reduce data label noise, and further improve the quality of labeling results.
[0047] This embodiment uses the power system data of a certain region in the past three years as the original data for scheme verification. The data involves two parts: power industry data and external data. Among them, the power industry data involves project contract master data, project execution process data and user evaluation data, etc.; external data involves meteorological data within three years, local economic benefit data, etc. In this paper, these data are randomly divided into four data sets to verify the proposed multi-source heterogeneous data fusion algorithm. The above four data sets contain 1,000 text instances, and each instance is described in different languages. In order to verify the performance of the algorithm described in this article, the control group algorithm is set to two algorithms: canonical correlation analysis and deep matching network. Both the experimental group and the control group use the same hardware configuration: Intel Core i7-7500U processor, 2.70GHz main frequency, 32GB memory configuration, and Matlab is used as the software platform.
[0048] Figure 4The results of the experimental group and two control group algorithms under four data sets as training data are shown. As can be seen from the figure, the accuracy of the scheme in this paper is better than that of canonical correlation analysis and deep matching network under four data sets, with average accuracy 8.32% and 11.7% higher respectively. This is because the multi-source heterogeneous data fusion algorithm described in this paper is essentially a deep semantic matching model for domain private networks and domain common networks. By combining canonical correlation analysis and deep learning networks, multi-layer nonlinear transformations can be used to mine multi-layer feature structures. The application of transfer learning can make up for the semantic errors of cross-domain data instances and realize knowledge transfer from the source domain to the target domain.
[0049] Figure 5 The effects of the number of hidden layers and neurons in the deep learning network on the semantic error are shown. It can be seen from the figure that with the increase of the number of hidden layers and neurons, the semantic error shows a rapid downward trend. It is worth noting that with the increase of the number of hidden layers, the semantic error decreases faster. And when the number of neurons is 50, the semantic error curve basically tends to be stable, which means that the result has converged. The increase in the number of hidden layers and neurons can help improve the multi-layer correlation matching effect and effectively compensate for the semantic deviation between heterogeneous modal data.
[0050] The above embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention.
Claims
1. A method for constructing an electric power knowledge graph based on heterogeneous data sources, characterized in that: The following steps are involved: S1: Collect multi-source heterogeneous data; S2: Classify the collected heterogeneous data and convert unstructured data text into structured data text; S3: Use CRF algorithm to perform Chinese word segmentation on structured data text; S4: Use canonical correlation analysis and deep neural network to perform layer-by-layer semantic matching on the obtained word segmentation data, and use the matched data to build an electric power knowledge graph.
2. According to claim 1, a method for constructing a power knowledge graph based on heterogeneous data sources is characterized in that: The collected heterogeneous data are divided into power grid data, user data and social data according to their sources, and divided into structured data and unstructured data according to their types. According to the professional terminology in related fields, semantic calculations are performed on the collected heterogeneous data, so as to classify the collected heterogeneous data.
3. A method for constructing a power knowledge graph based on heterogeneous data sources according to claim 2, characterized in that: The converted structured data text is in three levels: a header indicating the data type; an index indicating the data address; and a storage indicating the data content.
4. A method for constructing an electric power knowledge graph based on heterogeneous data sources according to claim 3, characterized in that: Using the CRF algorithm, combined with pre-labeled sample data, the words with the highest probability of appearing in the structured data text are output as word segmentation.
5. A method for constructing an electric power knowledge graph based on heterogeneous data sources according to claim 4, characterized in that: A deep belief network is used as the basic model of the deep neural network. The deep belief network is composed of multiple restricted Boltzmann machines. The network parameters are trained layer by layer in forward and reverse directions. Canonical correlation analysis is integrated into the deep belief network to participate in the correlation analysis of source domain data and target domain data.
6. A method for constructing an electric power knowledge graph based on heterogeneous data sources according to claim 2, characterized in that: Semantic computing includes the following: First, determine the relevant core words and core word-related word sets in different categories, distribute and calculate the sum of similarities between heterogeneous data and core words and core word-related word sets in different categories, classify heterogeneous data with a similarity sum greater than 0, and divide the heterogeneous data into the category with the largest similarity sum.
7. The method for constructing a power knowledge graph based on heterogeneous data sources according to claim 5 is characterized in that: Layer-by-layer semantic matching includes the following: using a stacked autoencoder to encode the source domain and the target domain, extracting cross-domain co-occurrence data through the encoding results to obtain hidden layer features; The canonical correlation analysis of the encoded source domain and target domain models is performed, and the maximum correlation coefficient matrix between the two data domains is obtained according to the co-occurrence data characteristics.
8. The method for constructing a power knowledge graph based on heterogeneous data sources according to claim 6 is characterized in that: During the back-propagation process of the deep belief network model, the maximum correlation coefficient matrix is used to fine-tune the domain network parameters.
9. A method for constructing a power knowledge graph based on heterogeneous data sources according to claim 1, 2, 3, 4 or 5, characterized in that: The matched data are filled into the electric power knowledge graph framework according to the association between the data to complete the construction of the electric power knowledge graph. The electric power knowledge graph framework is divided into multiple levels according to the types, and the nodes between each level are connected through relationships.
10. A method for constructing a power knowledge graph based on heterogeneous data sources according to claim 1, 2, 3, 4 or 5, characterized in that: The electric power knowledge graph has query and update functions. The instructions for querying information include query node name and query content. When the electric power knowledge graph receives the update function, it determines the update node and updates the information of the relevant nodes according to the received information text.
Citation Information
Patent Citations
Knowledge graph construction method and device
CN113434634A
