Power grid intelligent dispatching method based on on-duty system

Abstract

The application discloses a power grid intelligent dispatching method based on a duty system, comprising the following steps: S1, constructing an electric power operation corpus based on electric power knowledge graph technology; S2, generating an electric power dispatching corpus and an electric power vocabulary in combination with a dispatching voice translation file; S3, realizing dispatching voice recognition based on feature extraction, mode matching and model training; constructing a voice model based on a Wavenet network; S4, adopting a depth-first algorithm to analyze, push, call and update acquired electric power text information or voice information; S5, based on the electric power dispatching corpus and the electric power vocabulary, pre-processing electric power text information, and adopting a BERT model to process the electric power text information to obtain extracted text representation features and construct a text feature representation vector; S6, adopting a fault classification and preplan matching method to assist an electric power dispatching scene. The scheme improves the speed and accuracy of power grid fault disposal and improves the intelligent level of power grid dispatching.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid dispatch automation, and particularly relates to a power grid intelligent dispatch method based on a duty system. BACKGROUND

[0002] Under the current traditional dispatch mode, with the increase of working time, the understanding ability of the dispatcher in the dispatch center to text and voice will inevitably decrease with the mental fatigue, especially when facing text and voice information with unclear meaning or fuzzy sound quality. Meanwhile, when an emergency event such as a fault occurs, a large amount of information will be generated in a short time, and it is difficult for manual processing to quickly grasp and record the key points in the information, so as to make a quick decision. At present, there is no power grid intelligent dispatch method based on natural language processing and voice recognition to formulate strategies for emergency faults efficiently and accurately. SUMMARY

[0003] The present application aims to solve the shortcomings of the prior art that lack of quick and accurate response to emergency faults, and proposes a power grid intelligent dispatch method based on a duty system, which assists the power dispatching scene by means of natural language processing and voice recognition technology, improves the speed and accuracy of power grid fault disposal, and improves the intelligent level of power grid dispatching.

[0004] One technical solution provided in the embodiments of the present application is a power grid intelligent dispatch method based on a duty system, including the following steps:

[0005] S1, constructing an electric power operation corpus based on an electric power knowledge graph technology;

[0006] S2, generating an electric power dispatch vocabulary and an electric power word bank in combination with a dispatch voice translation file;

[0007] S3, realizing dispatch voice recognition based on feature extraction, pattern matching and model training; constructing a voice model based on a Wavenet network;

[0008] S4, using a depth-first algorithm to analyze, push, call and update the obtained electric power text information or voice information;

[0009] S5, preprocessing the electric power text information based on the electric power dispatch vocabulary and the electric power word bank, and using a BERT model to process the electric power text information to obtain extracted text representation features and construct a text feature representation vector;

[0010] S6, using a fault classification and preplan matching method to assist the electric power dispatching scene.

[0011] As a preferred, S1 includes the following steps:

[0012] Based on the power dispatching regulations, the power grid historical data original corpus, the power grid operation knowledge is constructed, and the power knowledge graph is constructed to optimize the decision knowledge of the power grid safety as the target;

[0013] An entity extraction method based on machine learning is used to realize automatic identification of named entities in the power knowledge graph.

[0014] A deep implicit relationship extraction method based on Markov logic network is used to realize extraction of the relationship between entities in the power knowledge graph.

[0015] An attribute extraction method based on rules and heuristic algorithms is used to automatically extract corresponding attribute names and attribute values from the original corpus to realize extraction of entity attributes.

[0016] As a preferred, in S3, the dispatching voice recognition is realized based on feature extraction, pattern matching and model training, including the following steps:

[0017] For telephone voice in the power dispatching process, data enhancement is realized by reducing voice quality through noise addition and changing sound frequency, and MFCC feature extraction method is used for pretreatment to obtain voice feature vectors representing voice message recognition information; WaveNet is used to identify MFCC features.

[0018] As a preferred, the steps of the MFCC feature extraction method are as follows:

[0019] Pre-emphasize the voice signal, frame it, window it and perform FFT Fourier transform;

[0020] Calculate the power spectrum, and convert the filtered output of the triangular band-pass filter to logarithmic form using the relationship between the mel domain and the linear frequency;

[0021] DCT transformation is performed to obtain MFCC coefficients;

[0022] Calculate N / 3 MFCC coefficients+N / 3 first-order difference parameters+N / 3 second-order difference parameters+frame energy to form N-dimensional MFCC features.

[0023] As a preferred, WaveNet is used to identify MFCC features, including the following steps:

[0024] Cut the voice signal by frame, extract the MFCC features of the voice signal, and obtain the voice feature vectors representing the voice message recognition information;

[0025] Input the voice feature vector into WaveNet to obtain the text label corresponding to the voice feature vector;

[0026] Calculate the CTC loss function of the sentence text corresponding to the voice feature vector;

[0027] updating parameters regarding the speech model.

[0028] As preferred, in S4, the following steps are included:

[0029] Information search: if the voice information is obtained, the voice recognition server is called to identify the voice information into text information, the rule information is obtained according to the text information, and the search is performed by using the depth-first algorithm according to the rule information, and the search result is sent to the text processor by the searcher;

[0030] Information processing: the text processor receives the search result sent by the searcher, classifies and processes the search result, and returns the search result to the external module; the search result is directly called for task, or is displayed in multiple dimensions in the form of power grid operation index automatic generation, key information automatic mapping, and attention information panoramic icon annotation;

[0031] Rule analysis: according to the input text information, semantic analysis and frequency analysis are performed, combined with power grid operation knowledge, power grid operation state and user search purpose, the rule with the highest matching degree is obtained from the rule library;

[0032] Rule input: the external module returns the user's final selected search result to the rule input module of the searcher, and the rule input module performs rule input, which is used to create and update the rule library.

[0033] As preferred, in S5, the preprocessing of the power text information includes Chinese word segmentation, stop word removal and normalization.

[0034] As preferred, in S6, the method of fault classification and preplan matching is used to assist the power dispatching scene, including the following steps:

[0035] The power grid fault identification method based on cost-sensitive learning and model adaptive selection fusion is used to realize fault classification and reduce the influence of class imbalance on event identification results;

[0036] The similar decision preplan query is realized based on the short text matching method, and the decision preplan information is returned to provide decision support for dispatchers.

[0037] As preferred, the power grid fault identification method includes the following steps:

[0038] The BERT model is used to convert the segmented power grid monitoring information into a high-dimensional vector, and the mean value of the alarm data sentence vector is obtained;

[0039] A deep learning model is constructed, including a Bi-LSTM network using a cross-entropy loss function, inputting fault samples for supervised training and parameter adjustment, and obtaining a first sub-model with good recognition rate for large sample fault categories.

[0040] Customize a multi-classification cost-sensitive loss function to replace the cross-entropy loss function in the model, increase the misclassification cost of small samples in the model training process, and obtain a second sub-model capable of accurately identifying small sample faults;

[0041] Each fault sample is input into the first sub-model and the second sub-model for discrimination, and then a model self-adaptive selection fusion method is used to integrate the discrimination results, so as to obtain a final event recognition result and output.

[0042] As preferred, the similar decision plan query is realized based on a short text matching method, decision plan information is returned, and decision support is provided for dispatch personnel, including the following steps:

[0043] The BERT model is used to convert the power grid fault information and the fault plan information after word segmentation into high-dimensional vectors;

[0044] The BERT model is used to calculate the plan matching result, and the three groups of plans with the highest matching scores are taken to be used for actual scheduling reference.

[0045] The power grid intelligent dispatching method based on the on-duty system of the application has the advantages that natural language processing and voice recognition technologies are used to assist power dispatching scenes, important information in unstructured data is extracted through natural language processing and voice recognition before power text information and voice messages are directly received by dispatching on-duty personnel, and the important information is displayed in a structured form. When a fault or an alarm occurs, a large amount of information generated in the system can be classified according to importance, similar information in historical messages can be matched, and a plan is given according to the processing method of the similar information to assist the dispatcher in decision-making. In this process, the generated dispatching log, alarm information and identity information are recorded, tampering is prevented, and subsequent review is facilitated.

[0046] The above summary of the application is only a summary of the technical solutions of the application, in order to more clearly understand the technical means of the application, the application can be implemented according to the content of the specification, and in order to make the above and other purposes, characteristics and advantages of the application more obvious and easy to understand, the specific embodiments of the application are described below. BRIEF DESCRIPTION OF DRAWINGS

[0047] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments made with reference to the accompanying drawings. The drawings are only for the purpose of illustrating preferred embodiments and are not considered to be limiting on the application. Moreover, the same reference numerals are used throughout the drawings to designate the same parts.

[0048] Figure 1A flowchart of the power grid intelligent dispatching method based on the on-duty system of the present application. DETAILED DESCRIPTION

[0049] To make the objectives, technical solutions, and advantages of the present application clearer, further detailed descriptions will be given below in conjunction with the drawings and examples. It should be understood that the specific embodiments described herein are merely the best mode of the present application, which are used to explain the present application and do not limit the protection scope of the present application. All other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present application.

[0050] Before the example embodiments are discussed in more detail, it should be mentioned that some example embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts depict the operations (or steps) as sequential processes, many of the operations (or steps) can be implemented in parallel, concurrently, or simultaneously. In addition, the order of the operations can be rearranged. The processes can be terminated when their operations are completed, but can also have additional steps not included in the drawings; the processes can correspond to methods, functions, procedures, subroutines, subprograms, etc.

[0051] Embodiment one: as shown in the figure, the power grid intelligent dispatching method based on the on-duty system comprises the following steps: Figure 1

[0052] S1, constructing a power operation corpus based on power knowledge graph technology.

[0053] In S1, the following steps are included:

[0054] Based on the power dispatching regulations and the original corpus of the power grid historical data, the power grid operation knowledge is constructed, and the power knowledge graph of the optimization decision knowledge with the power grid safety as the target is constructed.

[0055] The entity extraction method based on machine learning is used to realize the automatic identification of the named entities in the power knowledge graph.

[0056] The deep implicit relationship extraction method based on Markov logic network is used to realize the extraction of the relationship between the entities in the power knowledge graph.

[0057] The attribute extraction method based on rules and heuristic algorithms is used to automatically extract the corresponding attribute names and attribute values from the original corpus, and realize the extraction of the entity attributes.

[0058] S2, generating a power dispatching corpus and a power vocabulary in combination with the dispatching voice translation file.

[0059] ​Specifically, the personnel of the regulatory agencies at all levels search information through the internal telephone system, and through voice recording synchronization and intelligent voice analysis technology, the telephone recording is decomposed into short sentences that can be recognized by the system, the information is queried through the data access platform and returned to the query personnel through voice synthesis technology. The query content includes power and electricity, regulations, dispatching records, telephone recordings, personnel information, equipment information, notification processes and other types of information, which are replaced by the original mouse and keyboard query data through voice interaction.

[0060] S3, based on feature extraction, pattern matching, model training to realize dispatch voice recognition; construct a voice model based on Wavenet network.

[0061] In S3, based on feature extraction, pattern matching, model training to realize dispatch voice recognition, including the following steps: for the telephone voice in the power dispatching process, reduce the voice quality through noise adding and change the sound frequency to realize data enhancement, and pretreat through the MFCC feature extraction method to obtain the voice feature vector representing the speech message recognition information; use WaveNet to identify the MFCC feature.

[0062] The steps of the MFCC feature extraction method are as follows:

[0063] Pre-emphasize the voice signal, frame it, window it, and perform FFT Fourier transform;

[0064] Calculate the power spectrum, and convert the filtered output result to logarithmic form using the relationship between the mel domain and the linear frequency;

[0065] Perform DCT transformation to obtain MFCC coefficients;

[0066] Calculate N / 3 MFCC coefficients + N / 3 first-order difference parameters + N / 3 second-order difference parameters + frame energy to form N-dimensional MFCC features.

[0067] Using WaveNet to identify the MFCC feature, including the following steps:

[0068] Cut the voice signal by frame, extract the MFCC feature of the voice signal, and obtain the voice feature vector representing the speech message recognition information;

[0069] Input the voice feature vector into WaveNet to obtain the text label corresponding to the voice feature vector;

[0070] Calculate the CTC loss function of the sentence text corresponding to the voice feature vector;

[0071] Calculate the gradient update parameters of the voice model.

[0072] The most critical part of the network structure of WaveNet is the use of multi-layered dilated causal convolutional layers. Causal means that the output of the convolution only depends on the input up to the current position, i.e. no future features are used, which can be understood as shifting the position of the convolution forward. Dilated means that the convolution is performed in a skip fashion, which effectively expands the receptive field after multiple stacks, thus learning the dependencies between long sequences. The speech signal is a one-dimensional sequence after MFCC feature extraction, so the convolutional layer in WaveNet uses one-dimensional convolution. The number of feature maps of the last convolutional layer in WaveNet is the same as the size of the dictionary, and after Softmax processing, each small piece of MFCC can get the probability distribution over the entire dictionary.

[0073] The speech model based on the WaveNet network is constructed, comprising:

[0074] a. Use the dilated causal convolutional layer to model the long-time scale features of speech synthesis as the result of predicting the t-th point according to the first t-1 points of a sequence, so it can be used to predict the sample point value in the speech; the basic formula is as follows:

[0075]

[0076] The main component of the WaveNet network model is that each convolutional layer of this convolutional network performs convolution on the previous layer, the larger the convolution kernel and the more the layers, the stronger the time domain perception ability and the larger the perception range. In the generation process, after generating a point, the point is put into the last point of the input layer to continue iteration generation. Because the sampling rate of speech is high, the requirement for the perception range in the time domain is large, so the dilated convolution model is used.

[0077] Because the model only has causal convolution and no recursive connection, the model training speed is faster than RNN, especially for very long sequences; for causal convolution, there is a problem that many layers or large filters are needed to increase the receptive field of convolution. Therefore, the embodiment uses the expansion of the convolution by size arrangement to increase the receptive field. The dilated convolution is to make the filter applicable to an area larger than the length of the filter itself by skipping part of the input. It is equivalent to generating a larger filter from the original filter by adding zeros.

[0078] b. Use sotfmax to calculate the probability distribution of the output audio intensity. Because the original audio is stored as a sequence of 16-bit (per time step) integer values, the softmax layer needs to output 65536 probability values for each time step. In order to facilitate operation, μ-law companding transformation can be applied to convert it, reducing the number of output probabilities to 256, and the formula is

[0079] c. Using a gated activation unit as the activation function to model the conditional distribution: The method for modeling the probability with additional conditions in step a includes global and local methods, which can be selected according to the actual effect during specific development. For the speech synthesis task, h in the formula in step a can be considered as the text information to be converted. The global method accepts a separate additional input h, which affects the model output at all time points, which can be expressed as

[0080] S4, using a depth-first algorithm to parse, push, call and update the obtained power text information or voice information.

[0081] In S4, the following steps are included:

[0082] Information search: If the voice information is obtained, the voice recognition server is called to recognize the voice information into text information, the rule information is obtained according to the text information, and the depth-first algorithm is used for search according to the rule information, and the search result is sent to the text processor by the searcher;

[0083] Information processing: the text processor receives the search result sent by the searcher, classifies and processes the search result, and returns the search result to the external module; according to the search result, the task is called directly, or the search result is displayed in multiple dimensions in the form of power grid operation index automatic generation, key information automatic mapping, and attention information panoramic icon annotation;

[0084] Rule analysis: according to the input text information, semantic analysis and frequency analysis are carried out, combined with power grid operation knowledge, power grid operation state and user search purpose, the rule with the highest matching degree is obtained from the rule library;

[0085] Rule entry: the external module returns the user's final selected search result to the rule entry module of the searcher, and the rule entry module performs rule entry, which is used to create and update the rule library.

[0086] S5, based on the power dispatch vocabulary and the power word library, the power text information is preprocessed, and the BERT model is used to process the power text information to obtain the extracted text representation features, and the text feature representation vector is constructed.

[0087] The preprocessing of the power text information includes Chinese word segmentation, stop word removal and normalization object.

[0088] Chinese word segmentation refers to cutting a Chinese character sequence into individual words according to a specific algorithm. The embodiment uses a statistical word segmentation method. The statistical word segmentation method refers to performing word segmentation by performing statistics on the probability of the appearance of words and characters through a statistical model. The frequency of the appearance of adjacent characters and the credibility of words are directly related. A threshold is set, and when the frequency is higher than a certain value, it can be considered that a word is formed. Therefore, this method is also called dictionary-free word segmentation. In practical applications, it is generally used together with the string matching-based word segmentation method, which can automatically identify new words and eliminate ambiguities, and has the advantages of high speed and high efficiency.

[0089] De-stopping refers to removing some separated useless words to avoid their impact on the next link.

[0090] Normalization objects refer to normalizing different descriptions of the same object, such as “main transformer” and “main variable”, which can be integrated into one object.

[0091] For example, in November 2018, Google released a large-scale pre-training language model BERT (Bidirectional Encoder Representation from Transformers) based on bidirectional Transformer. The pre-training model can efficiently extract text information and be applied to various NLP tasks. This research has refreshed the current optimal performance record of 11 NLP tasks by using the pre-training model.

[0092] S6, using the method of fault classification and preplan matching to assist power dispatching scene.

[0093] The method of using fault classification and preplan matching to assist power dispatching scene includes the following steps:

[0094] The power grid fault identification method based on cost-sensitive learning and model adaptive selection fusion is used to realize fault classification and reduce the influence of class imbalance on event identification results.

[0095] The similar decision preplan query is realized based on the short text matching method, and the decision preplan information is returned to provide decision support for dispatchers.

[0096] Multi-classification cost-sensitive loss function: the traditional loss function sets the same misclassification weight for all class samples, so the loss of minority class samples is easily submerged. A cost-sensitive loss function FL suitable for multi-classification problems is constructed; the calculation formula of FL is as follows:

[0097]

[0098]

[0099] Wherein: m, n respectively represent the sample category number and the total number of samples; y ij , p ij respectively represent the true probability and the predicted probability of the i-th sample belonging to the j-th category; e [0, 1] represents the modulation factor; L represents the cross-entropy loss function; alpha t [0, 1] is a weight factor that can distinguish the misclassification cost of different categories of samples: the larger the sample size, the smaller the misclassification cost of this category of samples, and vice versa.

[0100]

[0101] Wherein: N j represents the number of j-th category samples.

[0102] FL is composed of two parts: the first part is the traditional cross-entropy loss function L, and the second part is the cost-sensitive loss value calculation considering the influence of class imbalance. The weight of the two is adjusted by the modulation factor e: the smaller e is, the larger the proportion of the latter is, and the higher the degree of distinction of FL to each category of samples. As a cost-sensitive loss function, FL sets different cost factors for each category of samples, increases the misclassification cost of minority class samples, and thus improves the recognition accuracy of minority class samples.

[0103] The power grid fault identification method comprises the following steps:

[0104] The BERT model is used to convert the segmented power grid monitoring information into a high-dimensional vector, the vector dimension is set to 300, and the mean value is obtained to obtain the alarm data sentence vector; a deep learning model is constructed, including a Bi-LSTM network using a cross-entropy loss function, inputting fault samples for supervised training and parameter adjustment, and obtaining a first sub-model with good recognition rate for large sample fault categories;

[0105] A multi-class cost-sensitive loss function is defined to replace the cross-entropy loss function in the model, increase the misclassification cost of small samples in the model training process, and obtain a second sub-model that can accurately identify small sample faults;

[0106] Each fault sample is input into the first sub-model and the second sub-model for identification, and then a model adaptive selection fusion method is used to integrate the identification results to obtain the final event identification result and output.

[0107] The similar decision plan query is realized based on the short text matching method, the decision plan information is returned, and the decision support is provided for the dispatch personnel, comprising the following steps:

[0108] The BERT model is used to convert the segmented power grid fault information and fault plan information into a high-dimensional vector, and the vector dimension is set to 300;

[0109] The preplan matching result is calculated using the BERT model, and the three groups of preplans with the highest matching scores are used as actual scheduling references.

[0110] Embodiment two

[0111] The embodiment two of the present application can provide a duty system.

[0112] The duty panoramic dashboard module: the top of the module displays the total number of monitoring devices, the number of current fault devices, the preplan usage rate, the prewarning accuracy rate, and the fault discovery and disposal rate; the display block is divided into power supply area monitoring conditions, fault discovery and disposal rate, fault quantity trend, fault classification proportion, fault handling conditions, and fault type proportion.

[0113] The power supply area monitoring conditions: according to the company management range of the logged-in user, the corresponding regional map is displayed, the map is divided into two tabs, and the global monitoring / key monitoring is clicked to complete the switching of the viewing data, the running conditions of the power plants, devices, and lines in the specified range are displayed, the global monitoring includes all power plant, device, and line information within the range, and the key monitoring includes the device information added by the intelligent prewarning and key monitoring block.

[0114] The fault discovery and disposal rate: according to the company level within the management range, the fault occurrence number and disposal rate of the subordinate companies are displayed in the province and city, the fault occurrence number and disposal rate of the subordinate power plants are displayed in the district and county company, and the fault number and disposal rate of the devices within the power station are displayed in the power station, the top ten are displayed, and the exceeding part is not displayed.

[0115] The fault quantity trend: the trend proportion of the fault occurrence under the user range is displayed, and the fault classification types are displayed respectively.

[0116] The fault classification proportion: the total number of fault occurrences under the user range is displayed, and the fault classification types are displayed in a pie shape; the fault classification is divided into: primary accident without stop, primary accident without stop, secondary accident without stop, secondary accident without stop, automation accident, communication accident, and other accidents.

[0117] The fault handling conditions: the fault handling trend under the user range is displayed.

[0118] The fault type proportion: the total number of fault occurrences under the user range is displayed, and the fault type types are displayed in a pie shape; the fault type is divided into: fault tripping, emergency shutdown, control removal, AC grounding, and other.

[0119] The intelligent prewarning module: including key detection and global detection.

[0120] Global monitoring includes all device information within the user management scope.

[0121] The screening conditions include the company, device name, voltage level, station, and device status, and the options are selected by drop-down selection, supporting fuzzy search.

[0122] In the table field, the device operating life: the first number is the operating life, and the second number is the operating life.

[0123] The device status is: normal, about to fail, and failure.

[0124] The on-duty information module includes fault component information, fault conversion information, fault extraction information, fault review information, and artificial intelligence training case library.

[0125] The screening conditions for fault component information, fault conversion information, and fault extraction information include the station, voltage level, and occurrence time.

[0126] The screening conditions for fault review information include the device name, device type, station, company, fault type, fault classification, voltage level, occurrence time, severity level, and review status.

[0127] The screening conditions for the artificial intelligence training case library include the device name, device type, station, company, fault type, fault classification, voltage level, occurrence time, review status, severity level, processing status, and processing time.

[0128] The account management module includes:

[0129] The account data is obtained from the platform through web crawler technology; the account is divided into company, power plant, and device levels.

[0130] Click the company name to view the company name, click the power plant to jump to the station page, and click the device to jump to the device phone.

[0131] The on-duty report module includes:

[0132] The on-duty report is divided into daily, monthly, and annual reports.

[0133] The screening conditions include the on-duty report, report date, company, and station.

[0134] The report type is switched by the on-duty report, and the report date is set to the current day in the daily report, the selector is set to day; the report date is set to the current month in the monthly report, the selector is set to month; the report date is set to the current year in the annual report, and the selector is set to year.

[0135] System management module: including menu configuration, message notification, role management, user management and audit log management five functions.

[0136] The above specific embodiments are the preferred embodiments of the power grid intelligent dispatching method based on the duty system of the present application, and do not limit the specific implementation range of the present application. The scope of the present application includes but is not limited to the specific embodiments. Any equivalent changes made in accordance with the shape and structure of the present application are within the protection scope of the present application.

Claims

1. A power grid intelligent dispatching method based on a duty system, characterized in that: Comprise the following steps: S1, based on power knowledge graph technology to build power operation corpus; S2, combined with dispatch voice translation file to generate power dispatch corpus and power word bank; S3, based on feature extraction, pattern matching, model training to realize dispatch voice recognition; construct the speech model based on Wavenet network; S4, using depth-first algorithm, the obtained power text information or voice information is analyzed, pushed, called and updated; S5, based on power dispatch corpus and power word bank, the power text information is pretreated, and the BERT model is used to process the power text information to obtain the extracted text representation characteristics, and the text feature representation vector is constructed; S6, using the method of fault classification and preplan matching to assist power dispatch scene; In S3, based on feature extraction, pattern matching, model training to realize dispatch voice recognition, comprising the following steps: For telephone voice in the process of power dispatch, the voice quality is reduced by adding noise and changing the sound frequency to realize data enhancement, and the MFCC feature extraction method is used for pretreatment to obtain the voice feature vector representing the recognition information of the voice message; Using WaveNet to identify MFCC features; In S4, comprising the following steps: Information search: if the voice information is obtained, the voice recognition server is called to identify the voice information into text information, the rule information is obtained according to the text information, the depth-first algorithm is used according to the rule information to search, and the searcher sends the search result to the text processor; Information processing: the text processor receives the search result sent by the searcher, classifies and processes the search result, and returns the search result to the external module; according to the search result, the task is called directly, or the search result is displayed in multiple dimensions in the form of power grid operation index automatic generation, key information automatic mapping, and attention information panoramic icon annotation; Rule analysis: according to the input text information, semantic analysis and frequency analysis are carried out, combined with power grid operation knowledge, power grid operation state and user search purpose, the rule with the highest matching degree is obtained from the rule library; Rule input: the external module returns the user's finally selected search result to the rule input module of the searcher, and the rule input module inputs the rule, which is used to create and update the rule library; In S6, the method of fault classification and preplan matching is used to assist power dispatch scene, comprising the following steps: Using the power grid fault identification method based on cost-sensitive learning and model adaptive selection fusion, the fault classification is realized, and the influence of class imbalance on event recognition result is reduced; Based on short text matching method, similar decision preplan query is realized, and decision preplan information is returned to provide decision support for dispatch personnel; The power grid fault identification method comprises the following steps: The BERT model is used to convert the segmented power grid monitoring information into a high-dimensional vector, and the mean value is obtained to obtain the alarm data sentence vector; a deep learning model is constructed, including a Bi-LSTM network using cross-entropy loss function, input fault sample for supervised training and parameter adjustment, and a first sub-model with good recognition rate for large sample fault categories is obtained; A multi-classification cost-sensitive loss function is customized to replace the cross-entropy loss function in the model, and the cost of misclassification of small samples in the model training process is increased to obtain a second sub-model capable of accurately identifying small sample faults; Each fault sample is input into the first sub-model and the second sub-model for identification, and then a model self-adaptive selection fusion method is used to integrate the identification results to obtain the final event identification result and output the result. A similar decision plan query is implemented based on a short text matching method, and decision plan information is returned to provide decision support for dispatch personnel, including the following steps: The BERT model is used to convert the segmented power grid fault information and fault plan information into high-dimensional vectors. The BERT model is used to calculate the plan matching results, and the three groups of plans with the highest matching scores are selected for actual dispatch reference.

2. The power grid intelligent dispatching method based on the duty system according to claim 1, wherein: In S1, the following steps are included: Based on the power dispatching regulations and the original corpus of historical power grid data, a power knowledge graph is constructed to build an optimized decision knowledge of power grid operation with the goal of power grid safety; An entity extraction method based on machine learning is used to automatically identify the named entities in the power knowledge graph; A deep implicit relationship extraction method based on Markov logic network is used to extract the relationships between entities in the power knowledge graph; A property extraction method based on rules and heuristic algorithms is used to automatically extract the corresponding property names and property values from the original corpus to extract the entity properties.

3. The power grid intelligent dispatching method based on the duty system according to claim 1, wherein: The MFCC feature extraction method includes the following steps: Pre-emphasize, frame, window, and FFT Fourier transform are performed on the voice signal; The power spectrum is calculated, and the obtained power spectrum is filtered through a triangular bandpass filter, and the filtered output is converted to a logarithmic form using the relationship between the mel domain and the linear frequency; DCT transformation is performed to obtain the MFCC coefficients; N / 3 MFCC coefficients + N / 3 first-order difference parameters + N / 3 second-order difference parameters + frame energy are calculated to form an N-dimensional MFCC feature.

4. The power grid intelligent dispatching method based on the duty system according to claim 1, wherein: WaveNet is used to recognize the MFCC features, including the following steps: The voice signal is cut into frames, and the MFCC features of the voice signal are extracted to obtain a voice feature vector representing the recognition information of the voice message; The voice feature vector is input into WaveNet to obtain the text label corresponding to the voice feature vector; The CTC loss function of the sentence text corresponding to the voice feature vector is calculated; The gradient update parameters of the voice model are calculated.

5. The power grid intelligent dispatching method based on the duty system according to claim 1, wherein: In S5, the preprocessing of the power text information includes Chinese word segmentation, stop word removal, and normalization.

Images