A top-down search method based on power map

Abstract

The present invention belongs to a retrieval method, and specifically relates to a top-down retrieval method based on an electric power graph. It includes the following steps: Step 1: input of search terms; input of search terms from the outside; Step 2: semantic interpretation of search terms; semantic interpretation of search terms; Step 3: retrieval; Step 4: sorting; Step 5: update information; update the database of previously semantically interpreted vocabulary, and update the electric power knowledge graph. The significant effect of the present invention is: the present application utilizes the system concept of the knowledge graph, so that the retrieval information can start from the root node first, and the effective retrieval results are obtained according to the amount of data retrieved and the corresponding weight calculation, and then it is determined whether the retrieval should be moved down to the lower-level technical branch, thereby ensuring that the retrieval process is fast and effective, and at the same time the retrieval results can also be fed back to the update of the knowledge graph, which may add new sub-graph branches, and may also add parallel graph branches.

Description

Technical Field

[0001] The present invention relates to a retrieval method, and in particular to a top-down retrieval method based on an electric power spectrum. Background Art

[0002] The power system, after entering the intelligent power management, is called the smart grid. It is built on the basis of an integrated, high-speed two-way communication network. Through the application of advanced sensing and measurement technology, advanced equipment technology, advanced control methods and advanced decision support system technology, it realizes the goals of reliable, safe, economical, efficient, environmentally friendly and safe use of the power grid. Its main features include self-healing, incentives and protection of users, defense against attacks, providing power quality that meets user needs, allowing access to various forms of power generation, starting the power market and optimizing and efficient operation of assets. However, the information collected by the smart grid is difficult to use because of its imperfect information structure and the independence of each piece of information.

[0003] There have been several years of experience in the application of knowledge graphs in power systems, but simple knowledge graphs cannot completely solve the technical barriers caused by independent data in power systems.

[0004] Therefore, it is necessary to establish a top-down retrieval method based on power maps. Summary of the invention

[0005] In view of the defects of the prior art, the present invention provides a top-down retrieval method based on power spectrum.

[0006] The present invention is implemented as follows: a top-down search method based on power spectrum, comprising the following steps:

[0007] Step 1: Enter the search term;

[0008] Input search terms from outside;

[0009] Step 2: Semantic interpretation of search terms

[0010] Conduct semantic interpretation of search terms;

[0011] Step 3: Search

[0012] Step 4: Sorting

[0013] Step 5: Update information

[0014] Update the database of previously semantically interpreted vocabulary and update the power knowledge graph.

[0015] In the above-mentioned top-down search method based on the power spectrum, the step 1 includes the following steps:

[0016] Search the search terms. The search database includes the established semantic database of power maps and the historical search term semantic interpretation database. After the search results are obtained, execute step 2.

[0017] In the above-mentioned top-down search method based on the power spectrum, the step 2 includes the following steps:

[0018] If the search term is a word in the power map, the search term is directly used to search, the search result is fed back, and the method ends.

[0019] If the search term is a word that has been semantically interpreted in the past, the search is directly performed using the results of the historical semantic interpretation, the search results are fed back, and the method is terminated. When the method is used for the first time, the data of the words that have been semantically interpreted in the past is 0, so the semantic interpretation of the search term is directly performed. In subsequent executions, the results of the semantic interpretation will be gradually added to the database. Therefore, in the second and subsequent search processes, it is necessary to query the words that have been semantically interpreted in the past.

[0020] If the search term is not a word that has been semantically interpreted before, semantic interpretation is performed and subsequent steps are executed.

[0021] In the above-mentioned top-down search method based on the power spectrum, the step 2 includes the following steps:

[0022] The semantic interpretation method includes the following:

[0023] (1) First, perform vocabulary decomposition

[0024] If there are special terms for power system in the search word or search phrase, firstly decompose the part of the words.

[0025] Then break down the verb in the remaining search terms or phrases.

[0026] The remaining words are broken down into words with similar meanings in the field of electricity.

[0027] If the search term cannot be decomposed into words with similar meanings in the power field, approximate semantic decomposition will be performed. After the decomposition is completed, the staff will determine whether it is correct and effective.

[0028] The above approximate semantic decomposition can be performed using existing technologies, such as setting a training set and a validation set in advance, and then using an artificial intelligence learning method to perform machine training, and the trained artificial intelligence is used for approximate semantic decomposition.

[0029] If the result of manual judgment of approximate semantic decomposition is valid, then execute the subsequent steps; if the judgment result is invalid, adjust the content of the training set and the verification set, re-train the artificial intelligence, and then perform approximate semantic decomposition again.

[0030] (2) Assigning weights

[0031] Assign a weight to each decomposed word according to the following rules:

[0032] If the decomposition is performed according to the above-mentioned vocabulary decomposition rules, the total weight of power nouns is 0.4, the total weight of power verbs is 0.3, and the total weight of other vocabulary is 0.3. That is, the sum of the weights of all decomposed power nouns is 0.4, the sum of the weights of all power verbs is 0.3, and the sum of the weights of other vocabulary is 0.3.

[0033] If the vocabulary decomposition results are determined by staff in the aforementioned steps, the weight of each vocabulary is also determined by the staff based on experience.

[0034] In the above-mentioned top-down search method based on power spectrum, the step three includes the following steps:

[0035] The results decomposed in step 2 are searched separately. The search starts from the top branch of the power knowledge graph and each branch is searched separately.

[0036] The number of search results is multiplied by the weight obtained in step 2 to obtain the weighted search results.

[0037] That is, if n decomposition results are obtained in step 2, the weight of each decomposition result is Q ij , the number of retrieval results for each decomposition result is P ij , where i represents the i-th decomposition result, 1≤i≤n, and j represents the j-th branch of the knowledge graph, then there is a weighted retrieval result for the i-th decomposition result

[0038] S ij =Q ij ×P ij

[0039] Then sum up all weighted search results to get the total weighted search result M j ,Right now

[0040]

[0041] In the above-mentioned top-down search method based on the power spectrum, the step 4 includes the following steps:

[0042] For all total weighted search results M j Sum and get M all , and then make the following judgment

[0043] max(M j )≥0.6M all

[0044] As described above, a top-down retrieval method based on power spectrum, wherein the method further includes step 4 of user information weighting. When this step is executed, the original steps 4 and 5 become steps 5 and 6 respectively.

[0045] The step 4 of user information weighting includes the following contents:

[0046] The dimensions and specific weights of user information weights will vary depending on the coverage of the power knowledge graph.

[0047] The user information weight setting method of the present application is as follows: the user weight dimension includes two dimensions: power quantity and power stability. The weight data of different users are determined according to the user characteristics. For the convenience of description, A is used for each user. k and B k They respectively represent the weight of the power quantity dimension and the power stability dimension of the kth user.

[0048] (2) User weighted calculation

[0049] First take C k =max{A k , B k}, that is, take A k and B k The larger one among

[0050] In the above-mentioned top-down search method based on the power spectrum, the step 5 of sorting includes the following steps:

[0051] M' in this step j The value of M′ is obtained according to the following principle. If step 4 is not performed, then j The value of M in step 3 j If step 4 is executed, then M′ j The value of is M′ calculated in step 4 j The value of

[0052] For all total weighted search results M′ j Sum and get M all , and then make the following judgment

[0053] max(M′ j )≥0.6M all

[0054] If the above formula holds, and at max(M′ j ) still has lower-level categories in the knowledge graph branch, repeat steps 3 to 4 in the lower-level technical categories.

[0055] If the above formula holds, and at max(M′ j ) belongs to a knowledge graph branch with no subordinate classification, output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) adds a technical graph branch under the knowledge graph branch to which it belongs;

[0056] If the above formula does not hold, then output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) is added in parallel with the knowledge graph branch to which it belongs.

[0057] The significant effect of the present invention is: this application utilizes the system concept of knowledge graph, so that the retrieval information can start from the root node first, and the effective retrieval results are obtained according to the amount of data retrieved and the corresponding weight calculation, and then it is determined whether the retrieval should be moved down to the lower-level technical branch, thereby ensuring that the retrieval process is fast and effective. At the same time, the retrieval results can also be fed back to the update of the knowledge graph, which may add new sub-graph branches as well as parallel graph branches. DETAILED DESCRIPTION

[0058] Before carrying out the method of the present application, a basic power map must first be established. The power map can be a map covering multiple dimensions such as power construction, operation and maintenance, monitoring, user management, etc., or it can be a map of a certain field, such as equipment operation and maintenance.

[0059] For the convenience of description, the subsequent descriptions of this application all use equipment operation and maintenance as the description object of the power map. Technical personnel in this field can apply this application to other power map fields based on the technical content and relevant background knowledge of this application.

[0060] This application decomposes the operation and maintenance of power equipment into four secondary dimensions: ledger, inspection, emergency repair, and informatization. Among them, inspection is divided into two tertiary dimensions of testing and maintenance, and informatization is divided into three dimensions of operation, data integration, and debugging integration.

[0061] On the premise that the above contents are confirmed, the method of this application is started.

[0062] A top-down search method based on power map includes the following steps:

[0063] Step 1: Enter search terms

[0064] Enter search terms from outside.

[0065] Search the search terms, and the search database includes the established semantic database of power maps and the historical search terms semantic interpretation database. After the search results are obtained, execute step 2.

[0066] Step 2: Semantic interpretation of search terms

[0067] If the search term is a word in the power map, the word search is directly used, the search result is fed back, and the method ends.

[0068] If the search term is a word that has been semantically interpreted in the past, the search is directly performed using the results of the historical semantic interpretation, the search results are fed back, and the method ends. When the method is used for the first time, the data of the words that have been semantically interpreted in the past is 0, so the semantic interpretation of the search term is directly performed, and the results of the semantic interpretation will be gradually added to the database in subsequent executions. Therefore, the query of the words that have been semantically interpreted in the past needs to be performed in the second and subsequent search processes.

[0069] If the search term is not a word that has been semantically interpreted before, semantic interpretation is performed and subsequent steps are executed.

[0070] The semantic interpretation of the present application may use any semantic interpretation method of the prior art, or may use the semantic interpretation method provided by the present application.

[0071] The semantic interpretation method of the present application includes the following contents.

[0072] (1) First, perform vocabulary decomposition

[0073] If there are special terms for the power system in the search words or search phrases, first decompose these words. For example, words such as electric meter, transformer, harmonic, etc. are all special terms for the power system.

[0074] Then decompose the verbs in the remaining search terms or search phrases. For example, words that express actions such as power outage, disconnect, and close are all verbs that need to be decomposed.

[0075] The remaining words are decomposed into words with similar meanings in the power field. For example, power data structure is decomposed into three words: power, data, and structured.

[0076] If the search terms cannot be decomposed into terms with similar meanings in the power field, approximate semantic decomposition will be performed. After the decomposition is completed, the staff will determine whether it is correct and valid.

[0077] The above-mentioned approximate semantic decomposition can be performed using existing technologies, such as setting a training set and a validation set in advance, and then using artificial intelligence learning methods to perform machine training, and the trained artificial intelligence is used for approximate semantic decomposition.

[0078] If the result of the manual judgment of the approximate semantic decomposition is valid, then execute the subsequent steps; if the judgment result is invalid, adjust the content of the training set and the verification set, re-perform artificial intelligence training, and then perform approximate semantic decomposition again.

[0079] Due to the characteristics of the power system, the number of proper nouns and special verbs related to electricity is relatively limited, and those skilled in the art can completely establish a proper word library to assist in the implementation of this step.

[0080] In practical situations, in most cases, the search terms or search phrases can be decomposed into a combination of several power nouns and power verbs.

[0081] (2) Assigning weights

[0082] Assign a weight to each decomposed word according to the following rules:

[0083] If the decomposition is performed according to the above-mentioned vocabulary decomposition rules, the total weight of power nouns is 0.4, the total weight of power verbs is 0.3, and the total weight of other words is 0.3. That is, the sum of the weights of all decomposed power nouns is 0.4, the sum of the weights of all power verbs is 0.3, and the sum of the weights of other words is 0.3. For example, in the above power data structure, the decomposed "power" is a noun exclusive to the power field, so it accounts for 0.4; "data" and "structured" are not exclusive nouns in the power field, and the sum of the weights of these two words is 0.3, and the weight of each word is 0.15.

[0084] If the vocabulary decomposition results are determined by staff in the aforementioned steps, the weight of each vocabulary is also determined by the staff based on experience.

[0085] Step 3: Search

[0086] The results decomposed in step 2 are searched separately. The search starts from the top-level branch of the power knowledge graph, and the search is performed on each branch separately. For example, the secondary dimensions of power equipment operation and maintenance include four secondary dimensions: ledger, inspection, emergency repair, and informatization. Therefore, the search is performed on the above four dimensions separately.

[0087] The number of search results is multiplied by the weight obtained in step 2 to obtain the weighted search result.

[0088] That is, if n decomposition results are obtained in step 2, the weight of each decomposition result is Q ij , the number of retrieval results for each decomposition result is P ij , where i represents the i-th decomposition result, 1≤i≤n, and j represents the j-th branch of the knowledge graph, then there is a weighted retrieval result for the i-th decomposition result

[0089] S ij =Qij ×P ij

[0090] Then sum up all weighted search results to get the total weighted search result M j ,Right now

[0091]

[0092] Step 4: User information weighting

[0093] This step includes the following:

[0094] (1) User information weight setting

[0095] The dimensions and specific weights of the user information weights will vary depending on the coverage of the power knowledge graph. Those skilled in the art may set them based on past experience or in accordance with the user information weight setting method provided in this application.

[0096] The user information weight setting method of the present application is that the user weight dimensions include two dimensions: electric power and electric power stability. The weight data of different users are determined according to the user characteristics. For example, if the user pays more attention to the power consumption, but has lower requirements for the stability of power harmonics, three-phase imbalance, etc., the weight of the electric power dimension is higher, and vice versa, the weight of the electric power stability dimension is higher.

[0097] For the convenience of description, A k and B k They respectively represent the weight of the power quantity dimension and the power stability dimension of the kth user.

[0098] (2) User weighted calculation

[0099] First take C k =max{A k , B k}, that is, take A k and B k The one with the larger value,

[0100] Then make

[0101] This step is an optional step, that is, if it is necessary to weight the information according to the user's attributes, then this step is performed; if it is not necessary to weight the information according to the user's attributes, then the subsequent steps are directly executed.

[0102] Step 5: Sorting

[0103] M' in this step j The value of M′ is obtained according to the following principle. If step 4 is not performed, then j The value of M in step 3j If step 4 is executed, then M′ j The value of is M′ calculated in step 4 j The numerical value of .

[0104] For all total weighted search results M′ j Sum and get M all , and then make the following judgment

[0105] max(M′ j )≥0.6M all

[0106] If the above formula holds, and at max(M′ j ) still has lower-level categories in the knowledge graph branch, repeat steps 3 to 4 in the lower-level technical categories.

[0107] If the above formula holds, and at max(M′ j ) belongs to a knowledge graph branch with no subordinate classification, output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) adds a technical graph branch under the knowledge graph branch to which it belongs;

[0108] If the above formula does not hold, then output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) is added in parallel with the knowledge graph branch to which it belongs.

[0109] Step 6: Update information

[0110] Update the database of previously semantically interpreted vocabulary and update the power knowledge graph.

[0111] If the personalized graph information weights are not recalculated according to the different properties and needs of users in the above step five, this step directly updates the power knowledge graph.

[0112] If the aforementioned step 5 recalculates the weight of the personalized graph information according to the different properties and needs of the user, this step will only update the power knowledge graph for the user, that is, the information of the power graph is uniquely bound to the user, and the knowledge graph will have a better match with the user. Moreover, as users continue to search and update, the personalized trend of the power knowledge graph will become more obvious, and eventually a power knowledge graph for different users will be obtained.

Claims

1. A top-down search method based on power map, characterized in that: The steps include: Step 1: Enter the search term; Input search terms from outside; Step 2: Semantic interpretation of search terms Conduct semantic interpretation of search terms; Step 3: Search Step 4: Sorting Step 5: Update information Update the database of previously semantically interpreted vocabulary and update the power knowledge graph; The step 2 comprises the following steps: The semantic interpretation method includes the following: (1) First, perform vocabulary decomposition If there are special terms for power system in the search word or search phrase, the special terms for power system are first decomposed. Then break down the verb in the remaining search terms or phrases. The remaining words are broken down into words with similar meanings in the field of electricity. If the search term cannot be decomposed into words with similar meanings in the power field, approximate semantic decomposition will be performed. After the decomposition is completed, the staff will determine whether it is correct and effective. The above approximate semantic decomposition is performed using existing technology. The training set and the validation set are set in advance. Then, the machine training is performed using the artificial intelligence learning method. The trained artificial intelligence is used for approximate semantic decomposition. If the result of manual judgment of approximate semantic decomposition is valid, then execute the subsequent steps; if the judgment result is invalid, adjust the content of the training set and the verification set, re-train the artificial intelligence, and then perform approximate semantic decomposition again. (2) Assigning weights Assign a weight to each decomposed word according to the following rules: If the decomposition is performed according to the above-mentioned vocabulary decomposition rules, the total weight of power nouns is 0.4, the total weight of power verbs is 0.3, and the total weight of other vocabulary is 0.

3. That is, the sum of the weights of all decomposed power nouns is 0.4, the sum of the weights of all power verbs is 0.3, and the sum of the weights of other vocabulary is 0.

3. If the vocabulary decomposition results are determined by the staff in the above steps, the weight of each vocabulary is also determined by the staff based on experience; The step three comprises the following steps: The results decomposed in step 2 are searched separately. The search starts from the top branch of the power knowledge graph and each branch is searched separately. The number of search results is multiplied by the weight obtained in step 2 to obtain the weighted search results. That is, if n decomposition results are obtained in step 2, the weight of each decomposition result is Qi j , the number of retrieval results for each decomposition result is P ij , where i represents the i-th decomposition result, 1≤i≤n, and j represents the j-th branch of the knowledge graph, then there is a weighted retrieval result for the i-th decomposition result S ij =Q ij ×P ij Then sum up all weighted search results to get the total weighted search result M j ,Right now The method further includes step 4 of weighting user information. When this step is executed, the original steps 4 and 5 become steps 5 and 6 respectively. The step 4 of user information weighting includes the following contents The dimensions and specific weights of user information weights will vary depending on the coverage of the power knowledge graph. The user information weight setting method of the present application is as follows: the user weight dimension includes two dimensions: power quantity and power stability. The weight data of different users are determined according to the user characteristics. For the convenience of description, A is used for each user. k and B k They represent the weight of the power quantity dimension and the power stability dimension of the kth user respectively; User weighted calculation First take C k =max{A k , B k }, that is, take A k and B k The one with the larger value, Then make The step 5 sorting includes the following steps: M' in this step j The value of M′ is obtained according to the following principle. If step 4 is not performed, then j The value of M in step 3 j If step 4 is executed, then M′ j The value of is M′ calculated in step 4 j The value of For all total weighted search results M′ j Sum and get M all , and then make the following judgment max(M′ j )≥0.6M all If the above formula holds, and at max(M′ j ) still has lower-level categories in the knowledge graph branch, repeat steps 3 to 4 in the lower-level technical categories. If the above formula holds, and at max(M′ j ) belongs to a knowledge graph branch with no subordinate classification, output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) adds a technical graph branch under the knowledge graph branch to which it belongs; If the above formula does not hold, then output max(M′ j ) belongs to the knowledge graph branch, and the staff will determine whether it is necessary to add max(M′ j ) is added in parallel with the knowledge graph branch to which it belongs.

2. A top-down search method based on power map according to claim 1, characterized in that: The step 1 comprises the following steps: Search the search terms. The search database includes the established semantic database of power maps and the historical search term semantic interpretation database. After the search results are obtained, execute step 2.

3. A top-down search method based on power spectrum as claimed in claim 2, characterized in that: The step 2 comprises the following steps: If the search term is a word in the power map, the search term is directly used to search, the search result is fed back, and the method ends. If the search term is a word that has been semantically interpreted in the past, the search is directly performed using the results of the historical semantic interpretation, the search results are fed back, and the method is terminated. When the method is used for the first time, the data of the words that have been semantically interpreted in the past is 0, so the semantic interpretation of the search term is directly performed. In subsequent executions, the results of the semantic interpretation will be gradually added to the database. Therefore, in the second and subsequent search processes, it is necessary to query the words that have been semantically interpreted in the past. If the search term is not a word that has been semantically interpreted before, semantic interpretation is performed and subsequent steps are executed.

Images