An alarm inducement judgment method for power dispatch automation operation process

Abstract

The application provides an alarm inducement research method for power dispatch automation operation process, comprising a text entity extraction module, a similar work order retrieval module based on a knowledge graph, a work order matching module based on similarity calculation and a graph improvement module, which extracts different entity features of text by constructing a BERT+BiLSTM+Resnet+CRF entity extraction model, retrieves top-n related alarm work orders based on the power dispatch automation alarm knowledge graph by using breadth-first search and time descending order sorting, and comprehensively considers TF-IDF, cosine similarity and Siamese network matching to match the most relevant work order, and maps the research information of the error operation leading to the alarm and the corresponding solution method and safety measures based on the work order; the application can obtain alarm summary related entities by mining potential entity features of the alarm summary, position the alarm entity by retrieving the historical power dispatch automation alarm knowledge graph, discover the potential work items leading to the alarm, and return the alarm causes, solution methods and solution measures corresponding to the work items.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid operation and maintenance, and in particular to an alarm cause judgment method for power dispatch automation operation process. BACKGROUND

[0002] The power system is one of the infrastructures supporting industrial production and daily life. Various alarm information may occur in the actual automation operation process. The traditional power automation alarm cause judgment relies on manual intervention, which is inefficient and easily affected by human factors. Based on the alarm knowledge graph in the field of power dispatch automation, various entities in the power system and the association relationship between them are modeled and stored. Through alarm entity extraction of alarm phenomenon summary, the alarm entity in the knowledge graph is retrieved, the alarm work item is traced back from the alarm entity, and the alarm cause, solution method and safety measures are mapped based on the work item, which can locate and judge the alarm cause of power dispatch automation operation, and provide auxiliary support for fault handling. SUMMARY

[0003] The present application proposes an alarm cause judgment method for power dispatch automation operation process, which can obtain alarm summary related entities by mining the potential entity features of the alarm summary, locate the alarm entity by retrieving the historical power dispatch automation alarm knowledge graph, and return the alarm cause, solution method and safety measures by excavating the potential work items that lead to the alarm, thereby providing auxiliary support for power dispatch automation operation alarm judgment.

[0004] The present application adopts the following technical solutions.

[0005] An alarm cause judgment method for power dispatch automation operation process, the method comprising a text entity extraction module, a knowledge graph-based similar work order retrieval module, a work order matching module based on similarity calculation, and a graph improvement module. The method extracts different entity features of the text by constructing a BERT+BiLSTM+Resnet+CRF entity extraction model, retrieves top-n related alarm work orders using breadth-first search and time descending order sorting based on the power dispatch automation alarm knowledge graph, and matches the most relevant work order by considering TF-IDF, cosine similarity and Siamese network, and maps the judgment information of the misoperation leading to the alarm and the corresponding solution method and safety measures based on the work order.

[0006] The method comprises the following steps:

[0007] Step S1, text entity feature extraction and recognition classification are performed on the power dispatch automation operation alarm information to obtain text entity objects related to power dispatch automation operation;

[0008] Step S2, based on the text entity object, the atlas keyword retrieval is carried out, and a high matching degree work order set is found out;

[0009] Step S3, the alarm information and the alarm summary of the matched work order are subjected to text similarity calculation, and the most similar historical work order record is obtained;

[0010] Step S4, the work item causing the alarm phenomenon in the most similar work order record is searched, and the work item and the corresponding solution method and safety measures are returned, and after the cause of the current alarm is found out, the alarm work order record is formed and imported into the knowledge graph.

[0011] In step S1, the text entity feature extraction and identification classification are carried out on the power dispatch automation operation alarm information, and the text entity object related to the power dispatch automation operation is obtained, which includes the following steps:

[0012] Step S1A1, the work order record, historical data and account information, and operation regulation manual generated in the process of power dispatch automation operation are taken as source data, and the power dispatch automation operation data set is constructed through data cleaning;

[0013] Step S1A2, a BERT+BiLSTM+Resnet+CRF entity extraction model is constructed, and the training is carried out based on the power dispatch automation operation data set.

[0014] Step S1A3, the model trained in S1A2 is used for entity identification of the power dispatch automation operation alarm information, and various entities involved in the alarm information are returned.

[0015] In step S1A1, in the data cleaning process, for text data, the python-docx library is used for text extraction; for pdf file data, the OCR technology is used for text extraction; for excel file table data, the pandas is used for text extraction, and all the extracted text data are subjected to text classification according to the text structure information, and the string cutting and deduplication are carried out according to the text length, so as to obtain the initial power dispatch automation operation data set;

[0016] In step S1A2, the BERT model construction formula is as follows:

[0017]

[0018] V=BERT([CLS],T1,...,T n ,[SEP])=(e1,...,e n )

[0019] In the formula, Q={Q1,Q2,...,Q n} represents the initial text sequence, and T={T1,T2,...,Tn} represents the initial text vector obtained by encoding. (e1,...,e n ) represents the vector after adding the separator before and after the initial text vector;

[0020] In step S1A2, the BiLSTM model is composed of a vector representation layer, a bidirectional LSTM layer, and a linear layer. The LSTM model is composed of an input word X t , a cell state C t , a hidden layer state h t , a forgetting gate f t , a memory gate i t , and an output gate o t . The calculation process of the LSTM includes forgetting the information in the cell state and memorizing the information at the new time, retaining useful information for the subsequent, discarding useless information, and outputting the hidden layer state h t at each time step. The output information of the forgetting gate, the memory gate, and the output gate is controlled by the hidden layer state h t-1 at the last time and the current input X t ;

[0021] In step S1A2, the specific formula of the Resnet model is as follows:

[0022] y l =h(x l )+F(x l ,W l )

[0023] x l+1 =f(y l )

[0024] In the formula, x l , x l+1 represent the input and output of the Lth residual unit respectively, f represents the ReLU function, h(x l ) represents the identity mapping of x l , and F represents the residual function.

[0025] In step S1A2, the specific formula of the CRF model is as follows:

[0026]

[0027] In the formula, Z(X) represents a normalization factor, f k (y i-1 ,y i ,X,i) is a feature function, which represents the value of the feature y i at the ith position given the input sequence x and the output sequence y, and λ k represents a weight parameter.

[0028] In step S2, the graph keyword retrieval is performed based on the text entity to find a set of work orders with high matching degree, and the specific steps include:

[0029] In step S2A1, a set of to-be-matched work orders Set is constructed, and the set of work orders that are more matched with the alarm information is stored by matching the alarm phenomenon description similar to the historical work order record.

[0030] In step S2A2, the number of hops h of the breadth-first search is set, the corresponding alarm object entity is retrieved in the graph based on the constructed power dispatching automation alarm knowledge graph, and the breadth-first search is started from each alarm object entity, and all work orders involved within the maximum number of hops are added to the to-be-matched work order set Set.

[0031] In step S2A3, the work orders in the to-be-matched work order set are sorted in descending order of time, and the top-n work orders on the time line are retained.

[0032] The power dispatching automation alarm knowledge graph constructed in step S2 includes but is not limited to the following entity types and relationship types:

[0033] The entities mainly include alarm work order record name, alarm phenomenon description, alarm reason, solution measure, and safety measure; the relationships mainly include triggering, solving, and sequence. The information corresponding to the entities is stored in the form of attributes, wherein the alarm work order record name includes two attributes of entity id and entity name, and the entity id is the unique identifier of the alarm work order record.

[0034] In step S2, the work orders in the to-be-matched work order set are sorted in descending order of time, and the top-n work orders on the time line are retained, so that the records in the to-be-matched work order set are highly related to the alarm reasons mentioned in the alarm summary. The top-n work orders on the time line are sorted according to the severity of the problem, the scope of influence, and the urgency of the solution, so that the retained work order records reflect the latest system status and problem trend.

[0035] In step S3, the alarm information and the alarm summary of the matched work order are subjected to text similarity calculation to obtain the most similar historical work order record, and the specific steps include:

[0036] In step S3A1, a weight set λ=(λ1, λ2, λ3) is set, which respectively represents the weight of the score after the text is calculated by TF-IDF, cosine similarity, and Siamese neural network.

[0037] In step S3A2, the TF-IDF value of the alarm information and the alarm summary of the matched work order is calculated to obtain the similarity S1.

[0038] Step S3A3, calculate the cosine similarity of the alarm information and the alarm summary of the matched work order, to obtain similarity S2;

[0039] Step S3A4, use the Siamese network to calculate the similarity of the alarm information and the alarm summary in the matched work order, to obtain similarity S3;

[0040] Step S3A5, obtain the similarity score S of the text through weighted calculation of the similarity;

[0041] Step S3A6, take the work order with the highest similarity score as the most matched work order.

[0042] In step S3A2, the TF-IDF similarity calculation formula is as follows:

[0043]

[0044] TF-IDF = TF·IDF

[0045] In the formula, n i,j represents the number of occurrences of the word t i in the document d j , TF i,j represents the frequency of the word t i in the document d j . |D| represents the number of all documents, and |j:t i ∈d j | represents the number of documents containing the word t i ;

[0046] In step S3A3, the specific formula of the cosine similarity is as follows:

[0047]

[0048] In the formula, A i , B i represent the values of each dimension of vectors A and B, represents the inner product of vectors A and B, represents the product of the modulus of A and B;

[0049] In step S3A4, the specific structure of the Siamese network is as follows:

[0050] The Siamese network is composed of two convolutional neural networks with the same structure, and the network structure is composed of convolutional layers, pooling layers, and fully connected layers. The convolutional layer is composed of a 3x3 convolutional kernel, and the pooling layer selects average pooling;

[0051] In step S3A5, the specific formula of the similarity weighted calculation is as follows:

[0052] S = λ1S1 + λ2S2 + λ3S3

[0053] In the formula, λ1, λ2, λ3 are weight parameters set in S3A1, S1, S2, S3 are similarities calculated in S3A2, S3A3, S3A4 steps.

[0054] In step S4, the work item causing the alarm phenomenon in the most similar work order record is searched, and the work item and the corresponding solution and safety measures are returned. After the cause of the current alarm is actually found out, a work order is formed to perfect the knowledge graph, which specifically includes the following steps:

[0055] Step S4A1, the most matching work order item obtained through step S3 is searched, the alarm part matching the current alarm phenomenon description in the alarm phenomenon content described in the work order item is searched, the corresponding reason is queried, and the alarm reason reason is obtained;

[0056] Step S4A2, the solution and the corresponding safety measures for the alarm reason reason in the most matching work item are searched, and the solution solution and the measures are returned;

[0057] Step S4A3, the alarm reason reason, the solution solution, and the safety measures measures are returned to the user in the form of a work ticket, providing auxiliary support for alarm information processing in the power dispatching automation operation process;

[0058] Step S4A4, the alarm reason finally found out and the corresponding solution, safety measures, etc. are formed into a work order, imported into the system, and the knowledge graph is perfected.

[0059] The method obtains alarm summary related entities by mining potential entity features summarized in the alarm information, positions the alarm entities by searching the historical power dispatching automation alarm knowledge graph, traces and explores potential work items causing the alarm, returns the alarm reason, solution measures, and safety measures corresponding to the potential work items, and forms alarm records from actual operation into the knowledge graph;

[0060] The text entity extraction module uses an entity extraction model to extract entities from the alarm text, so as to reduce human factor interference when extracting and analyzing the alarm summary generated in the power dispatching automation operation process;

[0061] The similar work order retrieval module based on the knowledge graph retrieves related entities based on the power dispatching automation alarm knowledge graph, and is used for quickly positioning the most relevant top-n work orders causing the alarm phenomenon;

[0062] The work order matching module based on similarity calculation is used to find the work order most matched with the alarm information, map the work item most likely to cause the alarm and the corresponding solution and safety measures from the work order, and provide auxiliary support for power dispatch automation operation alarm information positioning and research.

[0063] Compared with the prior art, the present application has the following beneficial effects: different entities in the alarm summary are better identified, the work order record most matched with the alarm summary is retrieved based on the power dispatch automation alarm knowledge graph, and the work item causing the alarm phenomenon and the corresponding solution and safety measures are mapped from the work order. BRIEF DESCRIPTION OF DRAWINGS

[0064] The application will be further described in detail below in combination with the drawings and specific embodiments:

[0065] ATTACHED Figure 1 is a schematic diagram of the principle framework of the present application. DETAILED DESCRIPTION

[0066] As shown in the figure, an alarm cause research method for power dispatch automation operation process, the method comprises a text entity extraction module, a similar work order retrieval module based on a knowledge graph, a work order matching module based on similarity calculation and a graph improvement module, which extracts different entity features of the text by constructing a BERT+BiLSTM+Resnet+CRF entity extraction model, retrieves top-n related alarm work orders using breadth-first search and time descending order sorting based on the power dispatch automation alarm knowledge graph, and comprehensively considers TF-IDF, cosine similarity and Siamese network to match the most relevant work order, and maps the research information of the misoperation causing the alarm and the corresponding solution and safety measures based on the work order.

[0067] The method comprises the following steps:

[0068] Step S1, text entity feature extraction and identification classification are performed on the power dispatch automation operation alarm information to obtain text entity objects related to power dispatch automation operation;

[0069] Step S2, graph keyword retrieval is performed based on the text entity objects to find a high matching degree work order set;

[0070] Step S3, text similarity calculation is performed on the alarm information and the alarm summary of the matching work order to obtain the most similar historical work order record;

[0071] Step S4, the work item causing the alarm phenomenon in the most similar work order record is retrieved, and the work item and the corresponding solution and safety measures are returned, and after the cause of the current alarm is actually found out, the alarm work order record is formed and imported into the knowledge graph.

[0072] In step S1, text entity feature extraction and identification classification are performed on the power dispatch automation operation alarm information to obtain text entity objects related to power dispatch automation operation, and the following steps are included:

[0073] In step S1A1, work order records, historical data, and account information generated in the process of power dispatch automation operation, and operation regulation manuals are taken as source data, and a power dispatch automation operation data set is constructed through data cleaning.

[0074] In step S1A2, a BERT+BiLSTM+Resnet+CRF entity extraction model is constructed, and training is performed based on the power dispatch automation operation data set.

[0075] In step S1A3, the model trained in S1A2 is used for entity recognition on the power dispatch automation operation alarm information, and various entities involved in the alarm information are returned.

[0076] In step S1A1, in the data cleaning process, for text data, the python-docx library is used for text extraction; for pdf file data, OCR technology is used for text extraction; for excel file table data, pandas is used for text extraction; all extracted text data is classified according to the structure information of the text, and string cutting and deduplication are performed according to the length of the text to obtain an initial power dispatch automation operation data set.

[0077] In step S1A2, the BERT model is constructed according to the following formula:

[0078]

[0079] V=BERT([CLS],T1,...,T n ,[SEP])=(e1,...,e n )

[0080] In the formula, Q={Q1,Q2,...,Q n} represents the initial text sequence, T={T1,T2,...,T n} represents the initial text vector obtained by encoding, and (e1,...,e n ) represents the vector after adding a separator before and after the initial text vector.

[0081] In step S1A2, the BiLSTM model is composed of a vector representation layer, a bidirectional LSTM layer, and a linear layer. The LSTM model is composed of an input word X t at time t, a cell state C t , a hidden layer state h t , a forgetting gate f t , and a memory gate it , output gate o t composition. The calculation process of LSTM includes: by forgetting the information in the cell state and the information at the new moment of memory, retaining the information useful for the subsequent, discarding the useless information, and outputting the hidden layer state h t at each time step t-1 and the current input X t control;

[0082] In step S1A2, the specific formula of the Resnet model is as follows:

[0083] y l =h(x l )+F(x l ,W l )

[0084] x l+1 =f(y l )

[0085] In the formula, x l , x l+1 respectively represent the input and output of the Lth residual unit, f represents the ReLU function, h(x l ) represents the identity mapping of x l , and F represents the residual function.

[0086] In step S1A2, the specific formula of the CRF model is as follows:

[0087]

[0088] In the formula, Z(X) represents a normalization factor, f k (y i-1 ,y i ,X,i) is a feature function, which represents the value of the i-th position feature y i under the given input sequence x and output sequence y, and λ k represents a weight parameter.

[0089] In step S2, the graph keyword retrieval is performed based on the text entity to find a high-matching degree work order set, and the specific steps include:

[0090] In step S2A1, a to-be-matched work order set Set is constructed, and the work order set matched with the alarm information is stored by matching the alarm information with the historical work order record similar alarm phenomenon description.

[0091] Step S2A2, set the hop number h of breadth-first search, based on the constructed power dispatch automation alarm knowledge graph, use the entities extracted in step S1A3 to retrieve the corresponding alarm object entity in the graph, start breadth-first search from each alarm object entity, and add all work orders involved within the maximum hop number to the set of work orders to be matched Set;

[0092] Step S2A3, sort the work orders in the set of work orders to be matched in descending order of time, and keep the top-n work orders in the time line.

[0093] The constructed power dispatch automation alarm knowledge graph described in step S2 includes but is not limited to the following entity types and relationship types:

[0094] The entities mainly include alarm work order record name, alarm phenomenon description, alarm reason, solution measure, and safety measure; the relationships mainly include triggering, solving, and sequence. The information corresponding to the entities is stored in the form of attributes, wherein the alarm work order record name includes two attributes of entity id and entity name, and the entity id is the unique identifier of the alarm work order record.

[0095] The step S2 described in the step S2, the work order in the set of work orders to be matched is sorted in descending order of time, and the top-n work orders in the time line are kept, so that the records in the work order to be matched are highly related to the alarm reasons mentioned in the alarm summary. The top-n work orders in the time line are sorted according to the severity of the problem, the scope of influence and the urgency of the solution, so that the kept work order records reflect the latest system state and problem trend;

[0096] In step S3, the alarm information and the alarm summary of the matched work order are calculated for text similarity, and the most similar historical work order record is obtained, which includes the following steps:

[0097] Step S3A1, set a weight set λ=(λ1, λ2, λ3), which respectively represents the weight of the score after the text is calculated by TF-IDF, cosine similarity and Siamese neural network;

[0098] Step S3A2, calculate the TF-IDF value of the alarm information and the alarm summary of the matched work order, and get the similarity S1;

[0099] Step S3A3, calculate the cosine similarity of the alarm information and the alarm summary of the matched work order, and get the similarity S2;

[0100] Step S3A4, use Siamese network to calculate the similarity of the alarm information and the alarm summary in the matched work order, and get the similarity S3;

[0101] Step S3A5, get the text similarity score S by weighted calculation of the similarity.

[0102] Step S3A6, take the highest similarity score of the work order as the most matched work order.

[0103] In step S3A2, the TF-IDF similarity calculation formula is as follows:

[0104]

[0105] TF-IDF = TF IDF

[0106] In the formula, n i,j represents the number of occurrences of the word t i in the document d j , TF i,j represents the frequency of the word t i in the document d j . |D| represents the number of all documents, |j:t i ∈d j | represents the number of documents containing the word t i ; in step S3A3, the specific formula of cosine similarity is as follows:

[0107]

[0108] In the formula, A i , B i represent the values of each dimension of vectors A and B, represents the inner product of vectors A and B, represents the product of A and B modules;

[0109] In step S3A4, the specific structure of the Siamese network is as follows:

[0110] The Siamese network is composed of two convolutional neural networks with the same structure, and the network structure is composed of convolutional layers, pooling layers, and fully connected layers. The convolutional layer is composed of a 3x3 convolutional kernel, and the pooling layer selects average pooling;

[0111] In step S3A5, the specific formula of similarity weighted calculation is as follows:

[0112] S = λ1S1 + λ2S2 + λ3S3

[0113] In the formula, λ1, λ2, λ3 are weight parameters set in S3A1, S1, S2, S3 are similarities calculated in steps S3A2, S3A3, S3A4.

[0114] In step S4, the work item causing the alarm phenomenon in the most similar work order record is searched, and the work item and the corresponding solution and safety measures are returned. After the cause of the current alarm is actually found out, the work order is formed to perfect the knowledge graph, which specifically includes the following steps:

[0115] In step S4A1, the most matching work order item obtained through step S3 is searched, the alarm part in the alarm phenomenon content described in the work order item that matches the description of the current alarm phenomenon is found, the corresponding cause is queried, and the alarm cause reason is obtained;

[0116] In step S4A2, the solution and the corresponding safety measures for the alarm cause reason in the most matching work item are searched, and the solution solution and the measures are returned;

[0117] In step S4A3, the alarm cause reason, the solution solution, and the safety measures measures are returned to the user in the form of a work ticket, and auxiliary support is provided for the alarm information processing in the power dispatching automation operation process;

[0118] In step S4A4, the alarm cause finally found out and the corresponding solution, safety measures, etc. are formed into a work order, imported into the system, and the knowledge graph is perfected.

[0119] The method obtains alarm summary related entities by mining potential entity features summarized in the alarm information, locates the alarm entities by searching the historical power dispatching automation alarm knowledge graph, discovers potential work items causing the alarm, returns the alarm causes, solution measures, and safety measures corresponding to the potential work items, and forms alarm records by actual operation into the knowledge graph;

[0120] The text entity extraction module uses an entity extraction model to extract entities from the alarm text, so as to reduce human factor interference when extracting and analyzing the alarm summary generated in the power dispatching automation operation process;

[0121] The similar work order retrieval module based on the knowledge graph retrieves related entities based on the power dispatching automation alarm knowledge graph, and is used for quickly locating the most relevant top-n work orders causing the alarm phenomenon;

[0122] The work order matching module based on similarity calculation is used to find the most matching work order with the alarm information, and the work item most likely causing the alarm and the corresponding solution and safety measures are mapped from the work order, thereby providing auxiliary support for the alarm information positioning and judgment in the power dispatching automation operation.

[0123] The embodiment comprises the following function modules running on a server cluster or a server: a text entity extraction module, a similar work order retrieval module based on a knowledge graph, a work order matching module based on similarity calculation, and a graph improvement module.

[0124] Among them, for the alarm summary generated in the power dispatch automation operation process, the traditional manual extraction and analysis method is easy to be affected by human factors, the entity extraction model is used for entity extraction of the alarm text, the human factor interference can be reduced, the related entity retrieval is carried out based on the power dispatch automation alarm knowledge graph, the most relevant top-n work orders causing the alarm phenomenon can be quickly located, the alarm text matching based on similarity calculation can find the work order most matched with the alarm, and the work item most likely causing the alarm, and the corresponding solution and safety measures are mapped out from the work order, so as to provide auxiliary support for alarm information positioning and judgment of the power dispatch automation operation.

[0125] The above is only a preferred embodiment of the present application, and is not intended to limit the present application in other forms. Any person skilled in the art can modify or change the above disclosed technical content to obtain equivalent embodiments. However, any simple modification, equivalent change and modification made to the above embodiments without departing from the technical solution of the present application, according to the technical essence of the present application, still belongs to the protection scope of the technical solution of the present application.

Claims

1. An alarm cause analysis method for power dispatch automation job process, characterized in that: Comprise the following steps; Step S1, text entity feature extraction and recognition classification are carried out on power dispatch automation operation alarm information, and text entity objects related to power dispatch automation operation are obtained; Step S2, graph keyword retrieval is carried out based on the text entity object, and a high matching degree work order set is found out; Step S3, text similarity calculation is carried out on the alarm information and the alarm summary of the matching work order, and the most similar historical work order record is obtained; Step S4, the work item causing the alarm phenomenon in the most similar historical work order record is searched, and the work item and the corresponding solution and safety measures are returned, and after the cause of the current alarm is found out, the alarm work order record is formed and imported into the knowledge graph; In step S1, text entity feature extraction and recognition classification are carried out on power dispatch automation operation alarm information, and text entity objects related to power dispatch automation operation are obtained, which specifically includes the following steps: In step S1A1, the work order record, historical data and account information, and operation regulation manual generated in the process of power dispatch automation operation are taken as source data, and the power dispatch automation operation data set is constructed through data cleaning; In step S1A2, a BERT+BiLSTM+Resnet+CRF entity extraction model is constructed, and training is carried out based on the power dispatch automation operation data set; In step S1A3, the model trained in S1A2 is used for entity recognition on the power dispatch automation operation alarm information, and various entities involved in the alarm information are returned; In step S2, graph keyword retrieval is carried out based on the text entity, and a high matching degree work order set is found out, which specifically includes the following steps: In step S2A1, a to-be-matched work order set Set is constructed, and the alarm information is matched with the alarm phenomenon description of the historical work order record, and the work order set matched with the alarm information is stored; In step S2A2, the number of hops h of breadth-first search is set, the corresponding alarm object entity is retrieved in the graph based on the constructed power dispatch automation alarm knowledge graph, and the breadth-first search is started from each alarm object entity, and all work orders involved within the maximum number of hops are added to the to-be-matched work order set Set; In step S2A3, the work orders in the to-be-matched work order set are sorted in descending order of time, and the top-n work orders in the time line are retained.

2. The alarm cause analysis method for power dispatch automation job process according to claim 1, characterized in that: In step S1A1, in the data cleaning process, the python-docx library is used for text extraction for text data; OCR technology is used for text extraction for pdf file data; and pandas is used for text extraction for excel file table data; all extracted text data is classified according to the structure information of the text, the string is cut and de-duplicated according to the length of the text, and the initial power dispatch automation operation data set is obtained; In step S1A2, the formula for constructing the BERT model is as follows: V = BERT([CLS], T1,..., T n , [SEP]) = (e1,..., e n ) In the formula, Q = {Q1, Q2, ..., Q} n } represents the initial text sequence, T = {T1, T2, ..., T} n } represents the initial text vector obtained through encoding; (e1,...,e n This represents the vector after adding delimiters before and after the initial text vector; In step S1A2, the BiLSTM model is composed of a vector representation layer, a bidirectional LSTM layer, and a linear layer; the LSTM model is composed of an input word X t , a cell state C t , a hidden layer state h t , a forget gate f t , a memory gate i t , and an output gate o t ; the calculation process of the LSTM includes forgetting the information in the cell state and memorizing the information at the new time, retaining useful information for the subsequent, discarding useless information, and outputting the hidden layer state h t at each time step t-1 , the output information of the forget gate, the memory gate, and the output gate is controlled by the hidden layer state h t of the last time and the current input X In step S1A2, the specific formula of the Resnet model is as follows: y l = h(x l ) + F(x l , W l ) x l+1 = f(y l ) In the formula, x l , x l+1 respectively represent the input and output of the Lth residual unit, f represents the ReLU function, h(x l ) represents the identity mapping of x l , and F represents the residual function; In step S1A2, the specific formula of the CRF model is as follows: where Z(X) represents a normalization factor, f k (y i-1 ,y i , X, i) is a feature function representing the value of the i-th position of the feature y i given the input sequence x and the output sequence y, and λ k represents a weight parameter.

3. The method of claim 1, wherein the method is characterized by: The constructed power dispatch automation alarm knowledge graph in step S2 includes, but is not limited to, the following entity types and relationship types: The entities mainly include alarm work order record name, alarm phenomenon description, alarm cause, solution measure, and safety measure; the relationships mainly include triggering, solving, and sequence; and the information corresponding to the entities is stored in the form of attributes, wherein the alarm work order record name includes two attributes of entity id and entity name, and the entity id is a unique identifier of the alarm work order record.

4. The method of claim 1, wherein the method is characterized by: In step S2, the work orders in the to-be-matched work order set are sorted in descending order of time, and top-n work orders in the time line are retained, so that the records in the to-be-matched work order are highly related to the alarm causes mentioned in the alarm summary; the top-n work orders in the time line are sorted according to the problem severity, influence range, and solution urgency, so that the retained work order records reflect the latest system state and problem trend; In step S3, the alarm information and the alarm summary of the matched work order are subjected to text similarity calculation to obtain the most similar historical work order record, which specifically includes the following steps: In step S3A1, a weight set λ=(λ1, λ2, λ3) is set, which respectively represents the weights of the scores obtained after the text is calculated by TF-IDF, cosine similarity, and Siamese neural network; In step S3A2, the TF-IDF value of the alarm information and the alarm summary of the matched work order is calculated to obtain a similarity S1; in step S3A3, the cosine similarity of the alarm information and the alarm summary of the matched work order is calculated to obtain a similarity S2; in step S3A4, the Siamese network is used to calculate the similarity of the alarm information and the alarm summary of the matched work order to obtain a similarity S3; In step S3A5, the similarity score S of the text is obtained through weighted calculation of the similarities; In step S3A6, the work order with the highest similarity score is taken as the most matched work order.

5. The method of claim 1, wherein the method is applied to power dispatch automation operation process. In step S3A2, the TF-IDF similarity calculation formula is as follows: TF-IDF=TF·IDF where n i,j represents the number of times the term t i occurs in the document d j TF i,j represents the frequency of the term t i in the document d j ; |D| represents the number of all documents, |j:t i ∈d j | represents the number of documents containing the term t i ; In step S3A3, the specific formula of the cosine similarity is as follows: where A i , B i denotes the value of the respective dimension of the vectors A, B, denotes the inner product of the vectors A and B, denotes the product of the modules of A and B; In step S3A4, the specific structure of the Siamese network is as follows: The Siamese network is composed of two convolutional neural networks with the same structure, and the network structure is composed of convolutional layers, pooling layers, and fully connected layers; the convolutional layers are composed of 3x3 convolutional kernels, and the pooling layers use average pooling; In step S3A5, the specific formula of the similarity weighted calculation is as follows: S=λ1S1+λ2S2+λ3S3 In the formula, λ1, λ2, and λ3 are weight parameters set in S3A1, and S1, S2, and S3 are similarities calculated in steps S3A2, S3A3, and S3A4.

6. The method of claim 1, wherein the method is applied to power dispatch automation operation process. In step S4, the work item causing the alarm phenomenon in the most similar historical work order record is retrieved, and the work item and the corresponding solution method and safety measure are returned; after the actual cause of the current alarm is found out, the work order is imported to perfect the knowledge graph, which specifically includes the following steps: Step S4A1, retrieve the most matched work order item obtained by step S3, find the alarm part in the content of the alarm phenomenon described in the work order item that matches the current alarm phenomenon description, query the corresponding reason, and obtain the alarm reason reason; Step S4A2, retrieve the solution and corresponding safety measures for the alarm reason reason in the most matched work item, and return the solution solution and safety measures measures; Step S4A3, return the alarm reason reason, solution solution and safety measures measures to the user in the form of a work ticket, and provide auxiliary support for the processing of alarm information in the power dispatching automation operation process; Step S4A4, form a work order for the final clear alarm reason and the corresponding solution and safety measures, import the system, and perfect the knowledge graph.

Images