Multi-scene process emergency system based on electric power emergency knowledge base and use method

Abstract

The invention relates to the technical field of electric power emergency management, in particular to a multi-scene process emergency system based on an electric power emergency knowledge base and based on natural language processing, a knowledge graph and deep learning and a use method. A traditional electric power emergency plan has the problems of low text structuring degree, insufficient dynamic response capability and the like, so that the emergency plan generation efficiency is low. In the prior art, the fusion capability of multi-source heterogeneous data is lacked, and plan disassembly depends on manual work, so that the requirement of quick response is difficult to meet. In addition, an existing system has obvious defects in the aspects of post responsibility matching, resource dynamic scheduling and the like. Based on the above technical scheme research conclusion, an emergency resource continuous optimization configuration method is analyzed, and a plan digital real-time information interaction technology is combined to realize integrated planning with a new-generation emergency command system to develop a demonstration application scheme.

Description

Technical Field

[0001] The present invention relates to the technical field of power emergency management, and specifically to a multi-scenario process-based emergency system based on a power emergency knowledge base based on natural language processing, knowledge graphs and deep learning, and a method for using the system. Background Art

[0002] Traditional power emergency plans suffer from poor text structure and insufficient dynamic response capabilities, resulting in inefficient emergency plan generation. Existing technologies lack the ability to integrate heterogeneous data from multiple sources, and plan decomposition relies on manual labor, making it difficult to meet the demand for rapid response. Furthermore, existing systems have significant shortcomings in job responsibility matching and dynamic resource scheduling. Summary of the Invention

[0003] The present invention proposes a multi-scenario process-based emergency system and its use method based on a power emergency knowledge base, which is implemented through the following technical solutions: A multi-scenario process-based emergency system based on the power emergency knowledge base. First, a digital plan technology framework is constructed, an emergency plan chapter event aggregation model is established, and a digital model and instruction template of the emergency plan are constructed; secondly, a structural decomposition of the plan is carried out, and the complex logical relationship representation and reasoning technology based on the emergency task set is analyzed to provide data support for the automatic extraction of key points in the emergency process and the adaptive construction of tasks; finally, a role-based multi-level power emergency knowledge association analysis is established for personnel with different roles, different scenarios, and different positions to realize the dynamic construction and generation of emergency response plans.

[0004] The method for using the multi-scenario process-based emergency system based on the power emergency knowledge base includes the following steps: (1) Build a digital emergency plan technology system, structurally decompose the power emergency plan through natural language processing and graph database technology, and establish an emergency plan chapter event aggregation model that includes event type, stage, link, and execution subject; (2) Extract semantic and structural features of the plan text based on graph neural network and BiLSTM-CRF model, generate chain task instruction templates, and build a power emergency knowledge graph; (3) Dynamically generate emergency response plans according to application scenarios, including: scenario element analysis based on knowledge elements and dynamic Bayesian network deduction of accident situation; generate emergency plans through dual similarity algorithm and cloud model decision-making; (4) Combine the job responsibility database with deep belief network DBN to generate job-oriented emergency response plans, and optimize instruction distribution through hybrid recommendation algorithm.

[0005] The specific structural disassembly in the step (1) includes: performing topic recognition, intelligent segmentation, clause completion, and dependency syntax analysis on the pre - plan text; extracting entity relationships through CRF and BiLSTM models to form emergency task instructions including trigger conditions, executors, and resource requirements.

[0006] The knowledge graph construction in the step (2) includes: integrating the set of emergency task instructions using the GridGraph graph engine; fusing the power emergency professional dictionary through entity disambiguation and linking technologies to generate a knowledge base containing more than 5 million entities and more than 1 million relationships.

[0007] The dynamically generating emergency response plan in the step (3) includes: matching historical scenarios based on the improved KNN algorithm of the case base; updating decision - making content through real - time data feedback and dynamically adjusting emergency resource requirements.

[0008] The emergency response plan generation in the step (4) includes: constructing a dynamic job responsibility database and integrating meteorological and power grid operation data; adopting the deep belief network DBN and collaborative filtering algorithm to achieve hybrid recommendation of instructions.

[0009] The present invention realizes the construction of a digital pre - plan technology system, integrates data such as laws and regulations and case bases, and establishes a power emergency command knowledge system; uses natural language processing technology to perform intelligent segmentation, clause separation, and dependency analysis on the pre - plan, extracts key entities (such as event types, execution subjects), and generates a chain - type task instruction template. Knowledge graph and dynamic plan generation, construct a power emergency knowledge graph based on the GridGraph graph engine to realize semantic association of task instructions; combine scenario deduction models and dynamic Bayesian networks to generate emergency response plans in real - time. Intelligent decision - making and resource scheduling, adopt deep reinforcement learning to optimize the pre - disaster pre - configuration and post - disaster dynamic scheduling of the four elements of "personnel - materials - equipment", and realize the hybrid recommendation of emergency response plans through DBN and improved KNN algorithms. Brief Description of the Drawings

[0010] Figure 1 Flow chart of the implementation of the invention patent; Figure 2 Schematic diagram of the partition of the emergency plan; Figure 3 Schematic diagram of the information included in typical instructions; Figure 4 Schematic diagram of typical instructions and their labels; Figure 5 Schematic diagram of the relationship between instruction - related labels; Figure 6 Schematic diagram of the calculation method of the matching degree of subject words; Figure 7 Schematic diagram of the intelligent segmentation logic of the emergency plan document; Figure 8 Syntactic dependency relationship schematic diagram; Figure 9 Schematic diagram of intelligent sentence splitting and sentence completion technical route; Figure 10 Schematic diagram of BiLSTM+CRF information extraction model; Figure 11 Text entity extraction schematic diagram; Figure 12 Schematic diagram of clustering feature tree structure; Figure 13 Knowledge element evolution graph construction process; Figure 14 Schematic diagram of emergency event scenario evolution path; Figure 15 Emergency plan optimization process based on cloud model. Specific implementation manners First, construct a digital pre-plan technical framework, establish an emergency plan chapter event aggregation model, construct a digital model of the emergency plan and an instruction template, and provide theoretical guidance for the research on pre-plan structured decomposition technology; second, conduct research on pre-plan structured decomposition technology, analyze the complex logical relationship representation and reasoning technology based on the emergency task set, realize the automatic extraction of key points in the emergency process and the adaptive construction of tasks, and provide data support for Technical Solution 3; then, for personnel in different roles, different scenarios, and different positions, research the role-based multi-level power emergency knowledge association analysis technology, and realize the dynamic construction of emergency response plans and the intelligent generation of emergency response plans.

[0012] Aiming at the problems that the existing power emergency plans cover long disposal processes, complex tasks and involve many departments, and the degree of structuring and digitization of the existing emergency plans is insufficient, combined with different power emergency scenarios, the emergency plans are classified and segmented, key knowledge and information are extracted, a set of digital plan technology systems and reconstruction models are established, and a method for constructing digital plans and a method for quickly generating emergency task instructions are proposed. First, the power emergency command knowledge system is sorted out. By sorting out materials such as laws and regulations, institutional standards, emergency plans, management regulations, operation manuals, and event cases, an emergency command knowledge system is formed; secondly, combined with digital technology, research is carried out on the disassembly technology, storage technology, and application technology of power emergency plans, and finally a digital emergency plan system is formed; then, research is carried out on the event aggregation model of the emergency plan chapters. Through the classification and sorting of documents, chapters, and events, the processing results of document classification, event merging, and chapter division are obtained. The event aggregation method is used for classification and grading and structure unification, and the chapter aggregation method is used for difference merging and feature selection, and finally an aggregation model is formed; then, the emergency tasks in the emergency plan are sorted out. Through expert disassembly, various types of information such as events, stages, and links are obtained, and through merging and optimization processing, functions such as marking, duplicate removal, complementation, and merging are realized; finally, according to the element correlation relationship, a chain task instruction template is formed, and finally a digital model of the emergency plan is constructed.

[0013] (1) Power Emergency Command Knowledge and Digital Power Emergency Plan Technology System 1) Establish a power emergency command knowledge system Sort out the relevant laws and regulations, institutional standards, emergency plans, management regulations, operation manuals, event cases, etc. of power emergency command, clarify the relevant regulations on emergency command in texts such as the Work Safety Law and the Emergency Response Law, and improve the source basis for the establishment of the emergency command knowledge system in combination with the specific requirements for emergency disposal of power industry emergencies; according to the division of responsible entities (government, production and operation units, emergency rescue teams) and event types (natural disasters, accident disasters, public health, social security), establish the emergency command system architecture; sort out the functions and application status of the existing power emergency command platforms and monitoring and early warning systems, and establish an emergency command technology support system.

[0014] 2) Establish a digital power emergency plan technology system Research on digital technologies based on theories such as natural language processing and graph databases, realize text processing, information extraction, storage and application of emergency plans, and establish a digital technology system for power emergency plans. Aiming at the characteristics of complex text structure and large content of emergency plans, research text preprocessing technologies, carry out research on automatic topic recognition, intelligent paragraph positioning, intelligent sentence splitting and completion, word merging and optimization technologies of plan texts, and realize the autonomous recognition, classification, disassembly and optimization of texts; on the basis of completing text preprocessing, use natural language processing to carry out key information annotation and extraction of sentence content, and use technologies such as knowledge graphs and neural networks to build the correlation between text information; use relational databases and graph database construction technologies to store the plan texts covering the correlation, which is convenient for subsequent query and application.

[0015] (2)Discourse Process Convergence Model Construct an event convergence model for emergency plan discourse, use the graph neural network algorithm and the BiLSTM-CRF sequence annotation method to construct a joint extraction model for emergency plan discourse-level events with multi-granularity semantic and structural information extraction, extract the key elements of the disassembled emergency plan text, calculate the target emergency plan main event vocabulary chain through semantic block division, summarize homogeneous event templates through clustering analysis, and automatically match and generate a digital disassembly logic framework for emergency plans to achieve "making the machine understand the plan"; 1) Sentence and Document Multi-Granularity Semantic Extraction and Representation Method Perceive the context globally from the emergency plan, and expand the receptive field to better identify trigger words and event arguments scattered in multiple sentences by obtaining semantic representations at different granularities (word granularity, sentence granularity, document granularity).

[0016] Convert sentence-level extraction into sequence annotation, and then splice the input and output of sentence-level extraction as the input of document-level extraction. Generally speaking, an event usually has a sentence in a text that can best express the event, that is, the central sentence of the event. From a semantic perspective, this sentence contains the most elements and information of the event, such as trigger words and more arguments. If the arguments can be supplemented on the basis of sentence recognition, the problem of argument dispersion will be solved. A DCFEE framework is constructed based on the method of "event central sentence + argument supplementation", and the event extraction process is divided into sentence-level extraction and document-level extraction. Sentence-level extraction uses the BiLSTM-CRF sequence annotation model, and the sentence is segmented into characters and input into the model to extract trigger words and arguments. Then, the input of sentence-level extraction is spliced with the obtained sequence annotation result as the input of document-level extraction. For the input sentence, judge whether it is the central sentence of the event, and on the basis of the central sentence, obtain the arguments in the sentences around this sentence in the document to supplement the event.

[0017] 2) Extraction of text passage meaning features and structural features Extract the sentence structure information (syntactic parsing tree) and semantic information (semantic parsing tree) in the text sequence, enhance the representation of the original sequence data, and learn and capture the unique features of the document data by modeling the graph structure, so as to obtain the correlation relationship of each element in the emergency plan.

[0018] From the perspective of structural features, use the graph structure to capture richer relationships between text elements for the best expression. First, convert the syntactic, semantic and other features of the original text into graph-structured data, and use the underlying structure information to solve it using the Graph Neural Network (GNN), such as constructing a syntactic dependency graph. The graph neural network is a learning framework based on message passing. By transforming, propagating and aggregating the features of nodes and edges, it can learn better graph representations and can model arbitrary graph-structured data. Secondly, perform graph representation learning, use a specially designed GNN to learn the unique features of different graph-structured data, transform event extraction into constructing a directed acyclic graph based on entities, first identify all event arguments in the document, and then solve the problem that the information does not flow across sentences caused by sentence-level extraction through document-level information fusion. Then, according to the set event type, judge whether the document contains an event of a certain event type. If so, gradually generate a directed acyclic graph of this event type.

[0019] 3) Cluster analysis of characteristic elements Fuse semantic features at different levels and structural features at different granularities for cluster analysis, obtain comprehensive and complete emergency plan information, and form a disassembled logical framework of the emergency plan divided by target. Among them, partition one includes the cover, approval page and signature page, partition two includes drawings, contact information tables, data submission tables, partition three includes general rules, risk and hazard degree analysis, and partition four includes prevention and early warning, information reporting, emergency response, post-disposal, plan management, and emergency support.

[0020] (3)Digitalization of emergency plans 1) Disassembly of text statements within partitions For various types of emergency plans of the company, based on theme recognition and plan partitioning, use the disassembly template to carry out disassembly work on the specific content of the plan to form task measure instructions.

[0021] On the basis of plan partitioning, considering that the text content of partition 4 (event grading, responsibilities of the organizational command structure, prevention and early warning, emergency response, information reporting, post-disposal, emergency support, plan management) is clear, the statements are mainly in the "subject-verb-object" structure and there are logical relationships, it is necessary to further refine the disassembly, refine the main body responsibilities, task measures, external conditions, etc., clarify the relationships between key contents, and form simple, clear, standardized and executable task measure instructions.

[0022] For example, in the original emergency plan text: "After the company's emergency response office receives the typhoon disaster warning information reported by each unit and the warning notice from the superior competent department, it immediately summarizes the relevant information, comprehensively analyzes the risks of facilities and equipment in conjunction with relevant functional departments, puts forward suggestions for the company's typhoon disaster warning, and issues the warning by the company's emergency response office after being approved by the company's special emergency leading group." From the original text of the plan, the information contained can be analyzed and divided into direct information and implicit information. The direct information includes: the executor (the company's emergency response office), task measures (summarize the relevant information, comprehensively analyze the risks of facilities and equipment in conjunction with relevant functional departments, put forward suggestions for the company's typhoon disaster warning, and issue the warning); the implicit information includes: the type of event (typhoon), stage (warning stage), link (warning release), scenario (issuing the warning), triggering conditions (after receiving the typhoon disaster warning information reported by each unit and the warning notice from the superior competent department), and prerequisite conditions (being approved by the company's special emergency leading group). Through the analysis of the information contained in the original text of the plan, the relevant attributes of the instruction are deduced (a total of 8 attributes are obtained).

[0023] 2) Construction of the chain task instruction template Through the analysis of the information contained in typical instructions, the relevant tags of typical instructions are summarized.

[0024] Through the analysis of the information contained in typical instructions, the relevant tags of typical instructions are summarized. Combining with the actual situation of emergency work, an instruction template covering the main instruction tags and their association relationships is constructed. The instruction template can be used to assist in the decomposition of text materials such as emergency plans, disposal plans, and emergency response plans, and can also be used to assist in the construction of an instruction library.

[0025] Specifically, it is an instruction of "simple, clear, standardized, and executable". The front and back parts are the attributes related to the association relationship of the instruction, including the "event, stage, link, scenario" corresponding to this instruction and the "triggering conditions, executor, recipient, prerequisite conditions, resource requirements" of this instruction, etc. The dark part in the front is the condition that must be met for the issuance of the instruction, that is, to meet "for what event, at what stage and link, in what situation, and the specific task measures to be executed". The light part in the back is the associated attributes of the task measures, that is, the "triggering conditions" of this task measure, the executor and recipient of the task measure, and the prerequisite conditions and resource requirements required to execute this task measure, etc.

[0026] In terms of plan decomposition, the digital emergency plan can provide support in scenarios such as the existing plan's structured decomposition and role configuration. Using a discrimination method that combines "semantic integrity + delimiter symbols", various types of emergency plans are decomposed using the instruction template to form a new digital plan based on instructions, which assists emergency duty officers in carrying out daily duty work, assists emergency management personnel in command and decision-making, and assists on-site personnel in emergency response.

[0027] 1) Pre - plan Disassembly For a single pre - plan, conduct theme recognition, paragraph and sentence disassembly. Through recognition annotation and relationship construction, form task measure instructions, including: ① Theme Recognition Using keyword matching technology based on a specific domain, judge the theme to which the document content belongs, and clarify the paragraph structure and text content to be disassembled. First, it is necessary to judge the theme to which the input pre - plan content belongs. After the theme is determined, the key text content to be parsed can be obtained according to the preset theme model. The theme recognition method proposed in this project is to use keyword matching technology based on a specific domain to identify the theme.

[0028] First, we will design a theme keyword model. Theme keywords are mainly used to match themes, and the specific information includes keywords and specific domains. For example, if the keyword is "typhoon" and the specific domain is "title", when the document is input, the system will obtain the document title domain and match the keyword. If the match is successful, the document will be classified into the defined theme. The purpose of distinguishing themes is that different themes correspond to different paragraph structures. After determining the theme, the sentences to be extracted can be extracted according to the corresponding paragraph structure, which can greatly improve the efficiency and accuracy of program operation.

[0029] The theme word matching degree is mainly matched through the file name, the title of the first page of the file, and the keywords in the file, and different weights are set. Among them, the file name has the highest weight, set to 2, the title of the first page of the file has a weight of 1.5, and the keywords in the file have a weight of 1. According to the weighted calculation, the matching degree value of the file and each theme word can be finally obtained. After using the theme recognition algorithm, the matching degree between the input file and the theme can be calculated, thus completing the theme recognition of the pre - plan.

[0030] ② Intelligent Paragraphing When a pre - plan document matches a theme word, it can correspond to the paragraph structure of the theme word. Since the computer program reads the document content as pure text data without paragraph structure information, it is necessary to use the method of text matching to match with the text data for text paragraphing. In the paragraph structure, we set information such as the starting text, ending text, paragraph number, and whether to extract. For the paragraphs that need to be extracted, the system automatically locates the starting text and ending text, and then takes the text content between the starting text and the ending text as the paragraph content.

[0031] In this project, the problem of page turning needs to be considered. That is, a paragraph may span multiple pages. Therefore, a loop structure needs to be set to continuously judge whether the subsequent page text contains the ending keyword. If it contains, the relevant text is intercepted; otherwise, continue to the next page.

[0032] Before performing intelligent segmentation, it is necessary to set the paragraph structure of the topic in advance. The paragraph structure adopts a tree structure, and the paragraphs to be extracted are marked, and the start keyword and end keyword of the extraction are set. The subtitle attribute structure of the document represents the paragraph structure of the document, and the subtitle field needs to be extracted. The link or stage to which the paragraph belongs needs to be set on the paragraph to be extracted. In this way, the text content obtained by intelligent segmentation automatically establishes an association relationship with the link or stage, and the items extracted in the text are all associated with the corresponding link or stage. With the help of paragraph positioning, not only can the text corresponding to each paragraph keyword be extracted, but also the default word of the current paragraph can be extracted. The default word can be used for the subsequent intelligent completion function to improve accuracy and efficiency.

[0033] ③ Intelligent sentence segmentation and sentence completion Intelligent sentence segmentation and completion involves segmenting paragraphs. This project intends to use a Chinese dependency parser, a core technology in natural language processing. Dependency parser aims to determine the syntactic structure of sentences by analyzing the dependencies between words within them. It analyzes the dependencies between words within a sentence (e.g., subject-verb relationships, meaning the relationship between subject and predicate, verb-object relationships, etc.), and automatically segments the sentence based on these dependencies and punctuation. After segmentation, it's necessary to determine whether the sentence contains parallel relationships or consecutive predicate structures. These sentences are generally related to the same topic, so the corresponding sentences are then merged to complete the missing parts.

[0034] The project team conducted comparative experiments on the use of jieba word segmentation to achieve intelligent sentence segmentation and completion. However, after conducting a large number of experimental tests on the text of the emergency plan, it was found that it was impossible to achieve the goal by only using jieba word segmentation to perform part-of-speech tagging, and then segmenting and completing sentences based on the part of speech of each word in the sentence. For example, there will be a situation where there is no object in the sentence, and it is meaningless to extract this sentence. At this time, jieba word segmentation cannot solve the problem. Therefore, on this basis, the Chinese dependency syntax analysis tool was introduced to determine whether the sentence structure is complete and whether the sentence content is meaningful. The subject in the sentence is then extracted through jieba word segmentation to replace the default word. In addition, for the emergency plan, the vocabulary of jieba word segmentation and Chinese dependency syntax analysis tools was modified to improve the accuracy, so that the goal of intelligent sentence segmentation can be better achieved, and the effect is basically the same as that of the sentences selected during manual annotation.

[0035] The significance of sentence completion is that the data used for model training is sentences that have been manually labeled with subjects, predicates, and objects. If the sentences are not completed, it will affect the effect of plan disassembly and extraction and reduce the accuracy.

[0036] The paragraph text is split by the intelligent sentence splitting and sentence completion module, resulting in several sentences. The syntactic dependency relationship of the sentences is determined to identify whether the subject or object needs to be supplemented.

[0037] The following table shows a sentence splitting result. It can be found that the paragraph is split into two sentences. According to the syntactic dependency relationship judgment, the second sentence is found to lack a subject, so the subject "company" is automatically added. But why is "company" added as the subject? It is because in the paragraph structure settings, the default subject for this paragraph is set as "company".

[0038] Table 1 Typical Task Instruction Tag Table ④ Information extraction The information extraction module mainly extracts the sentences formed after intelligent sentence splitting. The proposed solution for the project is to perform automatic extraction based on deep learning and machine learning models.

[0039] In terms of the selection of machine learning models, since information extraction based on sentences is a typical classification problem, we considered the Hidden Markov Model (HMM) and the Conditional Random Field Model (CRF).

[0040] In this project, through the training data, the model is estimated according to the maximum likelihood estimation method, and the state transition probability matrix is calculated. For example, if the number of times a certain tag appears as the first word tag in a sentence in the dataset is k, and the total number of sentences is N, then the probability of this tag as the first word of the sentence can be approximately estimated as k / N.

[0041] The Conditional Random Field (CRF) combines the characteristics of the maximum entropy model and the Hidden Markov Model. It is an undirected graph model and has achieved good results in sequence labeling tasks such as word segmentation, part-of-speech tagging, and named entity recognition in recent years. In this project, based on HMM, CRF can define more feature functions; by introducing custom feature functions, it can not only express the dependencies between observations but also represent the complex dependencies between the current observation and multiple previous and subsequent states. Therefore, the accuracy of CRF in this project is better than that of the HMM model, so the Conditional Random Field (CRF) is adopted as the machine learning model selected for this project.

[0042] In terms of deep learning models, we mainly considered extracting sentence features to construct feature engineering, and we compared various deep learning models.

[0043] Convolutional Neural Networks (CNNs) are a type of feedforward neural network with convolutional computations and a deep structure, and they are one of the representative algorithms of deep learning. Convolutional neural networks have the ability of representation learning and can perform shift-invariant classification on input information according to their hierarchical structure. Therefore, they are also known as "Shift-Invariant Artificial Neural Networks (SIANN)".

[0044] The research on convolutional neural networks began in the 1980s and 1990s. Time Delay Neural Network and LeNet-5 were the earliest convolutional neural networks. After the 21st century, with the proposal of deep learning theory and the improvement of numerical computing devices, convolutional neural networks have developed rapidly and have been applied in fields such as computer vision and natural language processing.

[0045] Convolutional neural networks are constructed by imitating the visual perception mechanism of organisms and can perform supervised learning and unsupervised learning. The sharing of convolutional kernel parameters in the hidden layer and the sparsity of inter-layer connections enable convolutional neural networks to learn grid-like features such as pixels and audio with a small amount of computation, have stable effects, and have no additional requirements for feature engineering of data.

[0046] The IDCNN model is a dilated convolutional network that removes the pooling layer in the CNN network, increases the dilated width, enlarges the receptive field without losing information, and at the same time allows each convolution to include a larger range of outputs.

[0047] Recurrent Neural Network (RNN) is a type of recursive neural network that takes sequence data as input, recurses in the evolution direction of the sequence, and all nodes (recurrent units) are connected in a chain.

[0048] The research on recurrent neural networks began in the 1980s and 1990s and developed into one of the deep learning algorithms in the early 21st century. Among them, bidirectional recurrent neural networks (Bidirectional RNN, Bi-RNN) and long short-term memory networks (Long Short-Term Memory networks, LSTM) are common recurrent neural networks.

[0049] Recurrent neural networks have memory, parameter sharing, and Turing completeness, so they have certain advantages in learning the non-linear features of sequences. Recurrent neural networks are applied in natural language processing (Natural Language Processing, NLP), such as speech recognition, language modeling, machine translation, etc., and are also used for various time series forecasting. Recurrent neural networks constructed by introducing convolutional neural networks can handle computer vision problems containing sequence inputs.

[0050] Long short-term memory networks are a special type of RNN that can learn long-term dependence information. LSTM has achieved quite great success in many problems and has been widely used. LSTM avoids the long-term dependence problem through deliberate design. Remembering long-term information is the default behavior of LSTM in practice, rather than an ability that requires a great cost to obtain.

[0051] All RNNs have a chained form of repeating neural network modules. In a standard RNN, this repeating module has a very simple structure, such as a tanh layer.

[0052] BiLSTM learns the semantic information between corpora through the parameters of the forget gate, memory gate, and output gate. It can capture the dependencies in the semantics in a bidirectional manner. Through the model, the score value of each word under each label can be output.

[0053] Through experimental calculations, in the information extraction of emergency plan statements in this project, the accuracy of the BiLSTM model is better than that of the IDCNN model. Therefore, the BiLSTM model is selected as the deep learning model for this project.

[0054] To sum up, this project selects the CRF model in machine learning and the BiLSTM model in deep learning as the information extraction models.

[0055] Based on BiLSTM, the transition probability matrix of the CRF layer can learn the constraints between labels from the finally predicted labels, reducing the number of invalid prediction tags. The final CRF layer can obtain the label with the highest selected prediction score as the final output.

[0056] After the model is selected, it needs to be trained according to the data characteristics, which are specifically divided into the following steps: First is data preprocessing. This step mainly converts the manually annotated data into "BIO" annotation. The "BIO" annotation mainly consists of the following tags: B-PER and I-PER represent the first character and non-first character of a person's name respectively B-LOC and I-LOC represent the first character and non-first character of a place name respectively B-ORG and I-ORG represent the first character and non-first character of an organization name respectively O indicates that the character does not belong to a part of the named entity B-WORK and I-WORK represent the first character and non-first character of the execution content respectively B-UP and I-UP represent the first character and non-first character of the superior department respectively B-DOWN and I-DOWN represent the first character and non-first character of the execution department respectively Through the data preprocessing stage, each character in the sentence can be "BIO" annotated.

[0057] Then is to train the model. The model often shows the phenomenon of overfitting during the training process. Therefore, it is necessary to clarify the data convergence interval. The usual method is to set the model loss function for calculation.

[0058] The score from the input X to the label y is: The probability from the input X to the label y is: The loss function of the model is: The purpose of model training is to make the loss function reach the minimum value, that is, the calculated probability reaches the maximum value.

[0059] In terms of model optimization, since the data in different fields have their own characteristics, it is often difficult to obtain good results by using existing models for training and prediction. Therefore, it is necessary to optimize the model parameters. The model parameter optimization strategy of this project mainly includes the following aspects.

[0060] One is to set the Dropout parameter to prevent overfitting. When not using Dropout to lose neurons, on a relatively small dataset, the network trains to fit this data, and the accuracy is close to 1, and the network shows the phenomenon of overfitting.

[0061] Therefore, set the dropout rate of Dropout to 0.5 to lose half of the neurons during the training process and prevent overfitting.

[0062] The second is to decay the learning rate in stages. If the learning rate is too small, your neural network won't be able to learn at all. If the learning rate is too large, overfitting is likely to occur. Therefore, adopt the method of decaying the learning rate in stages. Set the initial value of the learning rate to 0.001 and control it through the decay parameter of the optimizer class. At the beginning of the training, the learning rate is large to enable the model to learn quickly. Later, it continuously decreases to prevent overfitting, and the model can continue to learn and fine-tune in the later stage.

[0063] The third is to set mini-batch. Mini-batch is a batch of the training dataset at one time. If the batch gradient descent method is used, all the training sets need to be processed at once and then a gradient descent is implemented, which is very slow. While the mini-batch gradient descent method can divide the number of samples into multiple small mini-batches, so that individual small batches can be processed simultaneously to improve the learning speed, rather than processing all the X and Y training sets. Set this value to 128.

[0064] The fourth is to train multiple epochs (number of training rounds). When training, it is not enough to iterate all the data once. It needs to be repeated many times to fit and converge. If the number of rounds is too large, a large amount of useless resources will be consumed and the model won't be improved. If the number of rounds is too small, the training effect of the model is not good. Therefore, set it to 150 rounds, and the model reaches convergence and won't continue to train ineffectively.

[0065] In the above process, the basic process of constructing the information extraction model is given. Probability predictions are made for the "BIO" labels that each word in the sentence may belong to. The higher the score, the greater the possibility. Then, through the CRF transition probability matrix, the most likely label is finally selected as the output.

[0066] 2) Pre-plan zoning To accurately complete the decomposition of the emergency plan, conduct an overall analysis of the plan from the perspective of the text structure, and summarize that the emergency plan text mainly has four types of characteristics: Characteristic 1: The format and content are fixed, the form is standardized, and it is easy to disassemble. It corresponds to the cover, approval page, and signature page of the plan.

[0067] Characteristic 2: Chart structure, with a call relationship with the early warning and response work process parts. It corresponds to the attachments of the plan.

[0068] Characteristic 3: Simple form, belonging to principle-based descriptions, with weak relevance to other parts of the main text. It corresponds to the general rules and the analysis of risks and hazards degrees, two chapters of the plan.

[0069] Feature 4: Clear content, complete logic, and the sentences are in the structure of "subject + predicate + object" or its simplified structure. It corresponds to 8 chapters of the corresponding emergency plan, namely event classification, responsibilities of the organizational command structure, prevention and early warning, emergency response, information reporting, post-disaster handling, emergency support, and plan management.

[0070] Based on the actual types of emergencies and relevant emergency plans, on the basis of establishing the relationship between the instruction template and each tag, it is necessary to further segment, clause, and tokenize the plan text to extract the entities corresponding to specific event types, stages, links, scenarios, executors, trigger conditions, prerequisite conditions, recipients, task measures, resource requirements, etc. in the text. This involves computer natural language processing related technologies, including methods such as topic recognition, segmenting and clausing, semantic recognition, and obtaining additional information.

[0071] Fill the entities corresponding to the tags of event types, stages, links, scenarios, executors, trigger conditions, prerequisite conditions, recipients, task measures, resource requirements, etc. disassembled into the instruction template in sequence according to the corresponding tags and the text order, and number them according to the order of plan disassembly, marking "placeholder + linkage instruction identifier" to complete the entity filling of the instruction template.

[0072] In the actual emergency response process, it often involves the division of labor and cooperation among multiple departments and agencies, and the instructions also include the job responsibilities, activation conditions, etc. of different departments and roles. Therefore, in the process of plan digitization, it is necessary to clarify the logical relationship between each instruction. According to the dimension of link / role, use the links as column headings and all roles as row headings, and import the corresponding document instruction numbers in ascending order in the form of "document number - instruction number" into the corresponding positions; then delete the empty cells to form an instruction set with comprehensive content and implicit logical order. The corresponding number, the first digit refers to the document type number of the plan (operation instruction manual, response rule, emergency plan, on-site disposal plan, etc.), and the second digit refers to the task instruction number after the structured disassembly of the plan. For example, 3-4 refers to the task instruction "After the company's emergency response office receives the information reports from the leading department for event disposal and the incident unit, / the company's emergency response office / immediately verifies the nature, scope of influence, and losses of the event and reports to the company's deputy leader in charge" in the emergency plan. Based on this, the task instruction set considering the link and role is formed after sorting and integration.

[0073] Through the constructed instruction set, the targeted release of actual instructions can be supported. It can be accurate to the corresponding "links", corresponding "roles (department level)" and the release order of instructions corresponding to the same document. The release order of instructions is determined by combining the instruction numbers in the way of "first link, then role". Finally, after the instruction sets are collected, they are arranged in the order of priority: "1. Initial disposal link", "2. Analysis and judgment link", "3. Activation response link", "4. Initial information report". The instructions are arranged in an orderly manner according to the increasing of the instruction numbers of different role tasks in each link, and finally, in the form of an instruction flow, it assists the orderly development of on-site emergency work.

[0074] In terms of the representation and reasoning technology of the complex logical relationships of emergency tasks, by identifying the same content, removing duplicate task instructions and correcting text errors in the disassembled digital emergency plans, a special knowledge graph based on the GridGraph graph engine is constructed; then, through research on key entity extraction, new word discovery and other technologies, a professional dictionary for power emergency is constructed; finally, through entity disambiguation, entity linking and other technologies, knowledge fusion is achieved, and a power emergency knowledge base is constructed.

[0075] 1) Emergency task instruction set processing The processing of the emergency task instruction set includes three parts: same content recognition, task instruction deduplication and text error correction. It is the preliminary processing of the emergency task instruction set, and the processing result realizes the construction of the knowledge graph through GridGraph.

[0076] GridGraph is a power-specific graph database independently developed by the Institute of Intelligence Research. It is a database-class computing engine with independent intellectual property rights, supporting schema definition, multi-graph management, transaction capabilities, visualization, access control, and mixed data processing capabilities for relational data types, providing powerful core data management capabilities for business applications and enabling the "zero-knowledge" construction of power equipment knowledge graphs. At the same time, it provides the capabilities of large-scale data storage, analysis, and calculation in the power field, can provide the underlying data infrastructure for business systems in the power field, has dedicated high-speed processing for graph data, efficient graph analysis, and optimized processing algorithms, and can be applied to multiple fields, such as large-scale topology analysis and data management in fields such as power grid equipment and dispatching, traditional database replacement fields; it can also be applied to the localization and independent and controllable replacement of distributed graph databases; or applied to the field of large-scale power grid equipment management. GridGraph stores data with "nodes" and "edges" as the basic storage units to achieve the massive storage and parallel processing of data with complex association relationships; taking advantage of the physical "pointing" relationship between nodes, it can provide "index-free" association operations for adjacent nodes, having great technical advantages in applications based on network-type data structures; it is also a database specifically for processing "massive relationships", especially suitable for managing, analyzing, and processing application problems where there are interdependent relationships between a large number of objects. Therefore, for the task instruction set after pre-plan decomposition, a power emergency knowledge base can be constructed based on the GridGraph graph database engine.

[0077] GridGraph implements the Apache TinkerPop3 framework and is compatible with the Gremlin query language. Business application code can connect to the GridGraph graph database through multiple languages such as Java, Groovy, and Python to access its data.

[0078] The recognition of the same content during the construction process is achieved by combining edit distance and semantic similarity discrimination. First, for longer texts, they are initially segmented according to semantics or paragraph information. Then, the recognition of the same content for shorter text fragments is carried out.

[0079] By further analyzing the current emergency process status, using natural language processing technology for text intention understanding and entity extraction, performing fuzzy comparison of labeled entities based on the power emergency knowledge base, reasoning and extracting task key points, matching and positioning the labeled node tree-shaped emergency task instruction set, aggregating instruction content in combination with semantic rules, automatically constructing corresponding tasks, forming a digital emergency plan, and assisting in the development of various emergency work such as daily duty and command decision-making, it is mainly implemented by the following technical solutions.

[0080] 1) Semantic reasoning In common emergency response plans, the descriptions of specific support force deployments and emergency response operations generally correspond to strings with specific meanings. Based on the pre-planned keyword knowledge base, scanning and relevant rule matching are carried out on the basis of word segmentation and phrase recognition to identify semantic chunks with associated relationships, and describe the event parameters occurring during the operation process, such as inspection, air landing, and rescue parameters. Then, by associating the information of time and location with them, an action process ontology is expressed in the pattern of event, location, and time, so as to be able to monitor the status of key events and related entities and support the rapid triggering of different decisions for different events. The correlation analysis of emergency response operations can be processed according to the following rules: Emergency response operation := { start time < time, time reference >, involved location < geographical location >, force grouping < force deployment, support agency, support platform >, participants < force deployment, target object >, content < text description >...}. After the information parsing of the emergency response document is completed, the application scope of the candidate material information is matched to complete the processing and selection of the effective information, and an emergency response plan in a specified format is generated. Among them, for all contents except emergency task operations (such as material information such as situation reports and task descriptions), various semantic resources obtained are extracted according to the requirements of the emergency response decision-making template and can be directly filled into the corresponding data nodes in the plan. For the screening of materials in the process of emergency response operations, the inevitable time sequence relationship in the action events should be considered, and the conflicts existing in the business process relationship need to be resolved to facilitate the rapid adjustment of the application and disposal action plan. This study adopts the strategies of pre-constraint and post-constraint, and automatically matches according to the expected requirements of each action and the resource status. When the pre-constraint of an action is satisfied, the action node will be automatically selected as valid, and the post-constraint is that after processing this node, considering the impact on the status of other events and resource status, the expected status of each action node is dynamically corrected, and even the branch nodes are dynamically adjusted (including branch merging and the emergence of new branches), so that the adjustment of our emergency response operations can quickly adapt to complex situation changes.

[0081] 2) Semantic matching To address the semantic differences in the description of disasters in actual scenarios and the disaster ratings and severity levels in emergency documents, alignment at the semantic level is required. When encountering an actual disaster scenario, the actual scenario can be mapped to the relevant content in the emergency response plan to the greatest extent. At this time, a semantic matching model needs to be applied to automatically associate the required instructions according to the actual situation of the event.

[0082] Semantic matching models utilize similarity-based scoring functions. They measure the credibility of facts by matching the latent semantics of entities and the relationships contained in the vector space representation. The RESCAL (bilinear) model obtains its latent semantics by using a vector to represent each entity. Each relationship is represented as a matrix that models the pairwise interactions between latent factors. It defines the scoring function for a fact (h, r, t) as a bilinear function table. Where h and t represent the head and tail entities, and Mr represents the relationship matrix. This score can obtain the pairwise interactions between all components of h and t, and there are O(d2) parameters for each relationship. TATEC not only models the three-way interaction h = Mrt, but also models two-way interactions (such as the interaction between entities and relationships). The scoring function, where D is a diagonal matrix shared by all different relationships.

[0083] DistMult simplifies RESCAL by restricting Mr to a diagonal matrix. For each relationship r, it introduces a vector embedding r and requires Mr = diag(r). The scoring function only captures the pairwise interactions between the h and t components along the same dimension and reduces the number of parameters for each relationship to O(d). However, because for any h and t, hTdiag(r)t = tTdiag(r)h holds, this overly simplified model can only handle symmetric relationships, which is clearly not fully applicable to general KGs.

[0084] HolE combines the expressiveness of RESCAL with the efficiency and simplicity of DistMult. It represents both entities and relationships as vectors in Rd. Given a fact (h, r, t), first, the entity representations are combined using the circular correlation operation h * t ∈ R. Circular correlation compresses the pairwise interactions. Therefore, HolE only requires O(d) parameters for each relationship, which is more efficient than RESCAL. At the same time, because circular correlation is non-commutative, i.e., h * t is not equal to t * h. So HolE can model asymmetric relationships like RESCAL.

[0085] ComplEx extends DistMult by introducing complex-valued embeddings to better model asymmetric relationships. In ComplEx, the entity and relationship embeddings h, r, t no longer exist in the real space but in the complex space. The scoring function for defining a fact is: 3) Natural language generation Natural language generation (NLG) can be divided into two system architectures: streamlined and integrated. A streamlined natural language generation system is constructed by several modules that interact through input and output and are independent of each other. In an integrated natural language processing system, the modules interact with and cooperate with each other, which is more in line with the thinking process of the human brain. Deep learning is a commonly used method in NLG technology. It is a branch of the field of machine learning that learns and trains based on a sample library to build a model. In recent years, with the emergence of large-scale labeled data, deep learning technology has become a new research craze. The emergence of deep learning technology has solved the problem that traditional machine learning methods cannot obtain deep data features through shallow models. This outstanding feature extraction and selection ability has also been widely applied to natural language generation tasks.

[0086] In recent years, NLG technology based on deep learning has become a new hot topic. Deep learning technology can learn the internal laws and representation levels in sample data to obtain learning and analysis capabilities similar to those of humans. Deep learning requires a large amount of data. Constructing corpus data is an important link in building a deep learning text generation model. Each research field has its own data system and acquisition method. The acquisition of data affects the construction of the model, and data is often obtained through web crawling. Before text generation, the sentences in the corpus are first preprocessed and decomposed into word vectors through the distributed representation of words, so as to better represent the spatial logical relationship from words to vectors and facilitate the recognition and learning of the model.

[0087] For the adaptive construction task, it can be generated based on the predicted state or result of disaster losses. During this conversion process and the progress of the emergency response process, the requirements for emergency response are constantly changing, and the current situation is also changing. In this regard, this technical solution can adaptively construct instruction tasks by combining the implementation situation after extracting the key points of the emergency plan process, and can also support the targeted release of subsequent instructions.

[0088] Decompose the emergency plan system, propose a multi-level power emergency knowledge association analysis technology based on roles, clarify the duties and tasks of different command roles, conduct research on the information association of multi-role and multi-level emergency plans for different roles, different scenarios, and different positions of personnel, design a construction method for the entity model association diagram, and form a knowledge association topology sub-graph; for specific emergency scenarios, analyze the current situation and requirements of the emergency, clarify the emergency response tasks and executors, and update and push instructions in real time according to the feedback of the executors to achieve the dynamic construction of the emergency response plan. For the emergency response plan, first clarify the command role assumed by this position in different emergencies, analyze the emergency response responsibilities, and study the generation method of the emergency response plan based on the decomposition results of the digitalized plan and the power emergency knowledge base to achieve the intelligent generation of the emergency response plan for different positions.

[0089] Text clustering is widely used in text analysis and information retrieval. Initially, text clustering was only used to summarize texts. As the application fields of natural language processing have become increasingly rich, it has many new uses, such as improving search results, creating synonyms, and so on. Clustering algorithms can generally be divided into: partition-based, hierarchical-based, density-based, and model-based, represented by algorithms such as K-means, BIRCH, DBSCAN, and SOM respectively.

[0090] 2) Knowledge association analysis technology based on entity models Based on the proposed similar set of key features of instructions for instruction text clustering, it is also necessary to analyze the organizational structure relationship between these entity models. The association of entity models is the potential connection between entity models. Generally, there is no special connection between entity models until they are called. Model association can be achieved through various methods such as heuristic methods and framework methods. In this project, the association between entity models will be established through the association of business knowledge elements. In the emergency field, one business knowledge element can act on another business knowledge element, causing a change in its attribute state and generating corresponding outputs. Therefore, there is an input-output relationship between knowledge elements. Business knowledge elements can establish associations with each other in two ways: In the first case, when there are attributes with the same name among several business knowledge elements, the association can be established by comparing the attribute thresholds, that is, comparing the attribute values of these attributes with the same name. If the size relationship of the predetermined attribute values is met, these business knowledge elements will establish an association. In the second case, the attribute names of several business knowledge elements are different, but there can also be an association between these business knowledge elements. At this time, these business knowledge elements need to be associated through entity models.

[0091] There is an information model mapping between business knowledge elements and scenario data. The change in scenario data is reflected in the change of the attribute values of business knowledge elements, and some specific business knowledge elements are instantiated and activated. The change of one business knowledge element will also have an impact on other business knowledge elements through business knowledge element association and activate the corresponding business knowledge elements. In this way, a chain effect between business knowledge elements can gradually occur. This chain effect process between business knowledge elements can be represented by a business knowledge element evolution graph. Constructing a knowledge element evolution graph is an important tool for determining the associated model. The basic method of constructing a knowledge element evolution graph is to analyze the association between business knowledge elements on the basis of obtaining business knowledge elements, determine the type of association, then judge the knowledge element relationship method, find the appropriate successor knowledge element, and link the predecessor knowledge element and the successor knowledge element with a directed edge.

[0092] In this model association through the association method of knowledge elements, a chain of knowledge core - model - knowledge element - model can be formed. In entity model association, there are single - order relationships, And relationships, and Or relationships. If the entity model and the front - and - back business knowledge elements it links are regarded as a small system, then the association of business knowledge elements is the internal relationship of this system, while the association relationship between entity models is the external relationship between these small systems. For the And - type business knowledge element association through entity model association, the precursor business knowledge element, entity model, and successor business knowledge element can be regarded as an entity model system. For the Or - type knowledge element association through entity model association, multiple parallel entity model systems can actually be formed, that is, there are multiple paths from the precursor knowledge element to the successor knowledge element. The basic idea of obtaining entity model association through business knowledge element association is as follows: The And - type business knowledge element association represents one entity model, and the Or - type business knowledge element association represents multiple entity models; the successor business knowledge element of each entity model serves as the precursor business knowledge element of other entity models behind this knowledge element; taking the successor business knowledge element of an entity model as the precursor business knowledge element and using another entity model to associate these precursor business knowledge elements and other successor business knowledge elements.

[0093] The model association diagram can be regarded as a simplification of the knowledge element evolution diagram. By removing the business knowledge elements, only the entity models and the edges connecting the entity models remain, so the association situation of the models can be displayed more concisely. The method of transforming from the knowledge element evolution diagram to the entity model association diagram is as follows: Traverse the entire knowledge element evolution diagram in reverse order. Check the directed edge e that uses the entity model m to link the business knowledge element. If the precursor business knowledge element linked by this directed edge is linked by different model directed edges E, and each e'∈E cannot provide the input data required by e alone, then the entity model represented by E and the entity model represented by e are in an And relationship. If each e'∈E can provide the input data required by e alone, then the entity model represented by E and the entity model represented by e are in an Or relationship.

[0094] (2)Dynamic construction technology of emergency response plans based on application scenarios 1) Scenario construction of power emergency events Abstract the general laws of power emergency events occurring, and based on the current state of the event, reasonably analyze and infer the evolution path and future development trend of the disaster accident scenario, construct an effective evolution model, reveal the accident evolution mechanism, evolution path in power emergency events, and present the comprehensive situation.

[0095] First, based on the evolution process of power accidents, analyze the accident scenarios from five elements: the scenario state (S), emergency objectives (T), response measures (M), external environment (E), and the self-evolution of the disaster accident (D), so as to realize the expression of the evolution path of power accident scenarios, enabling the emergency decision-making entity to have an intuitive and general understanding of the current state of the sudden disaster accident, its possible future development trends, the factors affecting the development of the accident scenario, and the forces driving these factors to play their roles.

[0096] On this basis, then explore and analyze the evolution mechanism of power accidents and the evolution path of accident scenarios. Since there are multiple scenarios in the self-evolution process of a disaster accident under the influence of the external environment, for each scenario, the emergency decision-making entity will have different emergency objectives and response measures. Therefore, under the constraints of the external environment and self-evolution, when the disaster accident scenario reaches the emergency objective according to the response measures, at the next moment, the disaster accident will evolve to the desired scenario according to the wishes of the emergency decision-making entity; if it continues to be resolved reasonably, it will continue to evolve according to the expectations of the emergency decision-making entity, and the harm and development trend of the disaster accident will also be effectively controlled until the disaster accident disappears; when the scenario of the disaster accident does not reach the corresponding emergency objective according to the response measures, the disaster accident will evolve in a direction contrary to the intention of the emergency decision-making entity, and its harm and development trend may further deteriorate until the disaster accident is completely and reasonably resolved or the disaster accident naturally disappears. Based on this, a network structure diagram of the accident scenario can be constructed.

[0097] Finally, introduce the relevant theoretical knowledge of dynamic Bayesian networks, extract the key scenario state elements affecting the evolution process of the disaster accident, as well as the corresponding emergency objectives, response measures, external environment and other elements as the node variables of the dynamic Bayesian network, determine the value ranges of each node variable and the relationships between each node variable, and connect these node variables with causal relationships and changing over time to form a dynamic network of power disaster accident scenarios. Assign probabilities to the network node variables and calculate the state probabilities of each node variable, from which the direction and result of scenario evolution can be obtained, thus realizing the deduction of the disaster accident scenario.

[0098] Taking the general large - scale power outage event caused by the extreme convective disasters of strong wind and heavy rain affecting the Jiangsu power grid as an example for illustration. There are the following scenarios: The Jiangsu Meteorological Observatory issued a red alert for severe convective weather. The system meteorological forecast layer shows that in the next 24 hours, short - term heavy rainfall will occur in areas such as Nanjing, Suzhou, Yangzhou, Taizhou, Nantong, and Yancheng. The rainfall in some local areas can reach more than 220 mm, the average wind force reaches above level 12, and there is continuous strong lightning weather. It is expected that there is a possibility of tornado occurrence in some local areas. This severe convective weather has strong wind and heavy rain, and it is expected to have a greater impact on power grid equipment, possibly triggering events such as tower collapse and wire breakage, and backflow of rainwater into low - lying station buildings, which may seriously affect the safe operation of the power grid in the eastern coastal areas of Jiangsu Province. In extreme cases, there may even be a possibility of local large - scale power outages. At this time, according to the current state of the event, a reasonable analysis and reasoning can be carried out on the evolution path and future development trend of the event scenario, and an effective evolution model can be constructed. The specific operation path is as follows: From the start of the event outbreak, during the evolution process at time t1, two or more scenarios and two or more evolution paths appear, such as large - scale power outage, local power outage, or no power outage. Due to the intervention of the emergency decision - making subject, different emergency goals and disposal measures are formulated for the event scenario. While the event scenario evolves itself, it develops towards the expected emergency goal; whether or not the emergency goal is achieved, at time t2, the event scenario will evolve into another set of scenarios; and so on, until time t n At this time, the power supply is normal, and the following comprehensive situation evolution diagram of the scenario evolution can be constructed to provide decision - making reference for the emergency decision - making subject.

[0099] 2) Generation of emergency decision - making plans The "scenario - response" type of emergency method makes decisions based on scenario deduction, and the basis and foundation for response are scenarios. This part will take the results of power emergency event scenario deduction as the disposal goal, study the method of generating emergency decision - making plans based on knowledge element and case - based reasoning technology, and make dynamic adjustments and responses according to the changes in disaster accident scenarios, providing practical reference for scientific decision - making at the disaster accident site.

[0100] The extraction of scenario elements for power emergency events means regarding the emergency event as a combination of a series of interrelated scenarios. These interrelated scenarios are composed of different, independent scenario elements related to disaster accidents combined according to a certain logical relationship. Screening these different, independent scenario elements related to disaster accidents, and extracting the elements that can represent the accident state of a certain time segment. Specifically, the extraction content includes disaster - bearing elements, accident - source elements, disaster accident state elements, and environmental elements, which are used to describe the overall scenario. Secondly, based on the knowledge element theory, the attributes, states, etc. of these elements are expressed.

[0101] Next, the preliminary generation of the decision - making plan is carried out.

[0102] First, build an original case knowledge base to store historical power emergency event handling plans and their related evaluations, and conduct scenario-based expression and organization based on disaster-bearing elements, accident source elements, disaster accident state elements, and environmental elements. Construct corresponding indexing rules and establish a retrieval set. On this basis, construct a dual-similarity combination algorithm based on case scenario similarity and attribute similarity to calculate the global similarity of cases, serving scenario retrieval and scenario matching work.

[0103] Specifically, in terms of scenario similarity, use a text structure similarity algorithm with multi-feature fusion. First, build a thesaurus of topic words in the field of power emergency events and use the Word2vec model to perform word vector modeling for this specific field. Secondly, based on the Jaccard similarity algorithm of word vectors, calculate the lexical semantic similarity based on word vectors, the semantic shift similarity calculation, and the negation coefficient based on dependency syntax. Comprehensively consider the semantic similarity of words in the sentence, as well as the relative positions of words in the sentence and the problem of opposite scenarios. Next, perform multi-feature fusion of semantics through a BP neural network to achieve the calculation of scenario similarity. In terms of attribute similarity, according to the characteristics of power emergency events, divide the characteristic attributes into three categories: continuous numerical variables (such as building height, number of affected people, etc.), enumeration variables (such as high temperature, accident source state, etc.), and fuzzy variables (such as wind speed, precipitation, fire-fighting facilities, traffic smoothness, etc.). Adopt methods such as Euclidean distance calculation, direct assignment (for boolean values), and membership function algorithms to calculate the attribute similarity based on the nearest neighbor similarity algorithm. On the basis of calculating the two types of similarities, calculate the product of the two, that is, calculate the weighted similarity of attributes, so as to obtain the global similarity.

[0104] For the current emergency event, after completing the classification of the types of characteristic element variables, form a target library through hierarchical comparison. Then, the various characteristic elements of the target scenario can be compared with the scenarios in the case scenario library. Finally, achieve matching through the global similarity and list the set of similar historical cases.

[0105] Secondly, in order to generate decision-making solutions that are more adaptable to the current situation, it is necessary to evaluate and optimize the retrieved similar historical cases. The optimization of emergency plans is one of the key technologies for emergency decision-making in unconventional sudden disaster accidents based on "scenario-response". The key to emergency decision-making is to evaluate and optimize the generated alternative emergency plans. After generating relevant emergency plans based on the above methods, the next step is to optimize the alternative emergency plans. Aiming at the ambiguity and randomness problems existing in the optimization process of emergency plans for unconventional sudden disaster accidents, we propose an emergency plan optimization method based on the cloud model. The implementation process of this method mainly includes three parts: establishing a decision index set, evaluating emergency plans, and optimizing emergency plans. First, decision indexes should be selected and a decision index set should be established according to the principles of decision index selection and the actual situation of emergency plan formulation; then, based on the linguistic values of decision-makers and their corresponding number fields, the weight values and evaluation values of decision indexes are determined, and the digital characteristics of the cloud model are generated using the inverse cloud generator algorithm; then, using the idea of the comprehensive cloud and adopting cloud operation rules, the evaluation cloud of each emergency plan is obtained; finally, the evaluation cloud of the emergency plan is compared with the evaluation standard cloud of the emergency plan for similarity, the comment cloud and comments of each emergency plan are obtained, and the advantages and disadvantages of the emergency plan are comprehensively judged and the best emergency plan is optimized. Through the three digital characteristics of the cloud model, namely the expected value, entropy, and hyperentropy, this method integrates the ambiguity and randomness in the determination process of emergency decision-making index weight values and evaluation values, constitutes a mapping between qualitative and quantitative aspects, and can well complete the transformation between qualitative description and quantitative evaluation. In addition, it uses the group information of experts and uses qualitative language to describe the evaluation values and importance degrees of decision indexes, which conforms to the cognitive law and natural thinking of people, makes the decision-making result more intuitive and reasonable, and is convenient for giving play to the function of computer-aided decision-making.

[0106] Finally, due to the characteristics of uncertainty, derivativeness, and dynamics of emergencies, the uncertain multi-attribute decision-making information and the evolution of emergencies may both increase the risks in emergency response decision-making. Therefore, most emergency decisions are multi-attribute risk-based emergency decisions. Generally speaking, according to the real-time field information update and the feedback of plan evaluation, dynamically adjusting the numerical values and contents according to the existing plan structure can meet the overall emergency needs. However, in the context of complex power emergencies, more reliance is placed on decision-makers to revise and improve the overall plan, consider the evolution process of emergencies, dynamically revise the decision-making plan, and incorporate the entire event into the case knowledge base after handling risk events, so as to continuously learn and improve the decision-making ability. The mechanism introduction is as follows: First, according to the disaster situation data detected in real time, the classification results of the characteristic element variable types are transmitted in real time to the links such as case scenario matching, case screening, and decision plan optimization for real-time plan update and improvement.

[0107] Second, in complex scenarios, consider the dynamic adjustment of decision-making plans based on prospect theory (also known as the theory of bounded rationality). If the emergency plan initiated at time \(t_1\) cannot fully and effectively control the sudden scenario, and the trend of scenario evolution is gradually escalating and deteriorating, then at time \(t_{n + 1}\), the decision-maker timely and effectively adjusts the plan according to the latest information on the situation, minimizing the losses and impacts caused by the emergency. The profit and loss situation can be analyzed by constructing methods such as probability matrices and Bayesian networks. Further, the decision-maker directly gives the weights of the degrees of casualties and property losses. According to the idea of the weight function in prospect theory, the comprehensive value of the scenario is ranked to express the importance degree of the decision-maker to control the scenario under a certain plan. Finally, based on the above-mentioned comprehensive value of the scenario, the generated value of the plan, and the weight of the scenario, the expected prospect value of different plans is obtained. The above steps can help us use machines and decision-makers together to measure the prospects of each plan to select a certain plan at a certain time, thus forming the dynamic adjustment of the emergency decision-making plan.

[0108] Third, the completed disaster decisions are entered into the case knowledge base as historical cases, and new scenario indexing rules are continuously corrected and constructed. On the one hand, the historical case library can be enriched to provide assistance for future emergency event decisions; on the other hand, drawing on the idea of crisis learning theory, historical cases are evaluated and learned from to absorb lessons and continuously improve the emergency decision-making ability of decision-makers. In actual power emergencies, the factors causing power safety emergencies mainly include four aspects: one is the human factor. General human factors include cable theft, power theft, construction damage to electrical equipment, misuse and misjudgment by operators, etc. Special human factors include war and terrorist attacks, etc.; the second is the equipment factor; the third is the grid factor; the fourth is the natural factor, such as storms, floods, lightning strikes, fires, persistent fog, earthquakes, and ice cover, etc. However, these power emergencies basically ultimately result in power outages, and most power outages have experienced similar uncontrolled chain reactions, resulting in the power grid being split into islands.

[0109] 3) Research on the construction of an emergency resource command distribution model matching the application scenario and the disposal scenario By comprehensively using means such as on-site research, expert interviews, and analogical analysis of similar cases, clarify the types of emergency resource requirements for application scenarios and disposal scenarios, and construct an emergency resource instruction distribution model, which specifically involves the types, quantities, order of use, operation instructions, etc. of various emergency resources required in different emergency disposal scenarios, so as to facilitate the subsequent retrieval and prediction of emergency resource requirements. Build a communication network that meets the high-speed and reliable transmission of collected and interactive data, build a computer infrastructure that meets the smooth and stable operation of the system, and finally form an external hardware usage environment for the emergency resource requirement prediction model system integrating hardware devices such as integrated servers, workstations, communication network devices, video capture devices, and sensor devices. Based on the operation requirements of the emergency resource instruction distribution model system, draw on the micro-service structure design and ideas, design the support environment architecture; develop a load balancing service to improve the availability of the system under high concurrency; study the fast processing technology of demand data in a high-concurrency environment, study the storage and management technology of massive demand data, analyze the data characteristics and management requirements in the type library, quantity library, and usage method library of emergency resources, develop the corresponding database and its management system, and provide reliable storage and high-speed access services for various types of demand data.

[0110] Based on the constructed emergency resource instruction distribution model in different scenarios, use the Deep Belief Nets (DBN) to construct an emergency resource instruction distribution prediction model for various scenarios. The first layer of the DBN network is the input layer composed of signal source nodes, the second layer is the hidden layer, and the DBN neural network has multiple hidden layers. The data of the previous hidden layer is used to train the next hidden layer, that is, the DBN network is a neural network composed of multiple Restricted Boltzmann Machines (RBMs). The third layer is the output layer, and the output layer is a linear combination of the outputs of the hidden layer neurons. Take the emergency resource database as the input layer of the DBN neural network, which includes elements such as the categories, quantities, order of use, and usage methods of the required emergency resources; the second hidden layer is based on the emergency scenario type database, which includes elements such as the accident location, accident type, accident severity level division, and accident scenario description of various different emergency scenarios; the third layer, as the output layer, through the best combination and matching of the emergency resource element content of the input layer and the emergency scenario element information of the hidden layer in a linear combination manner, is manifested as the emergency resource instruction distribution output result based on various emergency disposal scenarios, providing guidance instructions such as the types, quantities, usage methods, and order of various guarantee resources required for on-site emergency disposal personnel to carry out emergency operations.

[0111] Taking the large - scale power outage event caused by the earthquake disaster affecting the Gansu power grid as an example, the demand analysis is carried out in combination with the power grid company's job responsibility database and the matched large - scale power outage emergency response plan, including the types, quantities, order of use, operation instructions and other specific contents of various emergency resources required in different emergency response scenarios. For example, the required personnel and their quantities, the specifications and quantities of utility poles, the quantities of high - voltage wires, etc. Finally, a Deep Belief Nets (DBN) is used to construct an emergency resource instruction distribution model for large - scale power outage scenarios.

[0112] By comprehensively using means such as on - site investigation, expert interviews, and analogical analysis of similar cases, the best - performing real - world cases in terms of application and disposal effects are selected to construct an application and accident disposal case library. Specifically, information such as the occurrence location, severity, disposal methods, required personnel position information, personnel quantities, and disposal methods of case accidents is collected as the data source for case - based reasoning of the KNN algorithm.

[0113] Since the improved KNN approach is a supervised learning algorithm based on instances and does not require training itself, only the appropriate parameter K needs to be selected for application. Therefore, in combination with the constructed accident case library, the cross - validation method is used to retrieve the accident case that best matches the sudden accident from the case library as the appropriate K value, extract the position information and personnel arrangement plan of emergency personnel in this case, calculate the distances between the sample to be classified and all learning samples, that is, compare and analyze with the resource prediction results of the emergency resource demand prediction model, extract the differences in personnel position information and resource requirements between the two, and improve and optimize the prediction results of emergency resources and the personnel information of accident cases as a supplementary strategy for tasks and measures in application scenarios and disposal scenarios under emergency conditions, enhancing its feasibility and operability when applied to real - world application scenarios and accident disposal scenarios.

[0114] The basic steps of case-based reasoning technology based on the improved KNN algorithm can be summarized into four main processes: case retrieval (Retrieve), case reuse (Reuse), case revision (Revise), and case retention (Retain). The problem to be solved or the emergency disposal situation is regarded as the target case, and the historical cases are called source cases. The set of source cases is called the case base. The basic process of this case recommendation technology to solve problems is as follows: taking the emergency disposal situation as the target case; using the description information such as the occurrence location and the severity of the accident of the target case to query past similar cases, that is, retrieving the case base to obtain source cases similar to the target case, and thus obtaining some solutions to the application scenarios and disposal scenarios in reality; if this solution fails, it will be adjusted to obtain a successful case that can be saved. After this process ends, a relatively complete solution to the target case can be obtained; if the source case fails to give a correct and appropriate solution, a new source case can be obtained through case revision and retention.

[0115] Taking the large-scale power outage event caused by the impact of extreme strong convective disasters such as strong winds and heavy rains on the Jiangsu power grid as an example, first, the large-scale power outage is regarded as the target case; then, using the description information such as the occurrence location and the severity of the accident of the target case to query past similar cases. For example, the occurrence location is Jiangsu, the accident is that the power grid is affected by extreme strong convective disasters such as strong winds and heavy rains, and the severity of the accident is serious. Query similar cases based on this to obtain source cases similar to the target case, and thus obtain some solutions to the application scenarios and disposal scenarios in reality, and finally achieve the matching of the application scenario and the disposal scenario.

[0116] 5) Research on the emergency disposal plan push technology based on the hybrid recommendation algorithm The emergency disposal plan push technology based on the hybrid recommendation algorithm is divided into two implementation stages: in the first stage, the content recommendation algorithm is used to determine the specific content information of the emergency disposal plan, and in the second stage, the collaborative filtering algorithm is used to screen the recommended content of the content recommendation algorithm to determine the final information such as the emergency positions and corresponding content of the emergency disposal plan.

[0117] In the first stage, according to the resource demand content predicted by the Deep Belief Network (DBN), using the content recommendation algorithm, relevant content in the emergency response plan library is extracted; based on the emergency position and the corresponding personnel arrangement plan output by the improved KNN algorithm, using the content recommendation algorithm, relevant content in the emergency response plan library is extracted. According to the resource information such as the type of resources and the use of resources input by on-site emergency response personnel in the DBN neural network prediction resource model, and the accident scenario information input in the KNN algorithm case reasoning model, the semantic analysis method LAS is used, and the document-word matrix singular value decomposition method is used to map the document and words into the same low-dimensional latent semantic space, so as to more accurately calculate the similarity between the preference profile and the project information file, thereby improving the recommendation performance; calculate the similarity between the resource information and accident scenario information input by the emergency response personnel and the existing information, and recommend the similar information of the historically browsed resources to them as the targeted emergency response plan content recommendation. Aiming at the problem of the user cold start dilemma caused by the limited historical behavior of emergency response personnel, a user adaptive fine-tuning strategy is designed in the pre-training-fine-tuning paradigm solution. During the fine-tuning process, the pre-trained model parameters are adaptively fine-tuned according to the input information of different emergency response personnel, and finally good recommendation performance is shown, alleviating the accurate recommendation problem in the cold start scenario.

[0118] In the second stage, the collaborative filtering recommendation algorithm is used to screen the result set of the first stage. The recommended content of the content recommendation algorithm in the first stage is processed into feature data. According to the information retrieved by the emergency response personnel, and calculate the predicted score of the person for the neighbor information with a greater similarity to this information. Then, the predicted scores are arranged in descending order, and the obtained data set is used as the final recommended information list for the emergency response personnel. The algorithms in the two stages finally form a hybrid recommendation algorithm to determine the specific emergency positions and corresponding content in the emergency response plan, and accurately push the emergency response plan.

[0119] Taking the large-scale power outage event caused by the impact of debris flow disasters on the Gansu power grid as an example, in the scenario of large-scale power outages, first, a content recommendation algorithm is used to determine the specific content information of the emergency response plan. For example, the command institutions for handling large-scale power outages are company leaders, assistant chief engineers, leading departments, and participating departments; the composition and responsibilities of the command headquarters; the response process; the work requirements of the incident unit; the requirements for the on-site arrival of staff; the work requirements for video emergency consultations; the contact information of government departments, etc. Then, a collaborative filtering algorithm is used to screen the recommended content of the content recommendation algorithm to determine the final information such as the emergency positions and corresponding content of the emergency response plan. Finally, the specific emergency positions and corresponding content on the emergency response plan are determined, and the emergency response plan is accurately pushed. For example, the emergency response plan is accurately pushed to terminals such as mobile phones or computers using mobile phone numbers or the accounts of different personnel. Thus, different personnel can make accurate, timely, and rapid responses to sudden large-scale power outage events.

Claims

1. A multi-scenario process-based emergency system based on the power emergency knowledge base, characterized by: First, a digital emergency plan technical framework is constructed, an event aggregation model for emergency plan chapters is established, and a digital emergency plan model and instruction template are constructed; Secondly, we conduct structural disassembly of emergency plans, analyze the complex logical relationship representation and reasoning technology based on emergency task sets, and provide data support for the automatic extraction of key points in the emergency process and the adaptive construction of tasks; Finally, for personnel with different roles, different scenarios, and different positions, a role-based multi-level power emergency knowledge correlation analysis is established to realize the dynamic construction and generation of emergency response plans.

2. The usage method of the multi-scenario process-based emergency system based on the power emergency knowledge base according to claim 1, characterized in that The following steps are involved: (1) Build a digital emergency plan technology system, structurally decompose the power emergency plan through natural language processing and graph database technology, and establish an emergency plan chapter event aggregation model that includes event types, stages, links, and execution entities; (2) Extract semantic and structural features of the plan text based on the graph neural network and BiLSTM-CRF model, generate a chain task instruction template, and build a power emergency knowledge graph; (3) Dynamically generate emergency response plans based on application scenarios, including: scenario factor analysis based on knowledge elements and dynamic Bayesian network deduction of accident situations; generate emergency response plans through dual similarity algorithm and cloud model decision-making; (4) Combining the job responsibility database with the deep belief network (DBN), we generate job-specific emergency response plans and optimize instruction distribution through a hybrid recommendation algorithm.

3. The usage method of the multi-scenario process-based emergency system based on the power emergency knowledge base according to claim 1, characterized in that: The structured decomposition in step (1) specifically includes: subject identification, intelligent segmentation, sentence completion and dependency syntax analysis of the plan text; extracting entity relationships through CRF and BiLSTM models to form emergency task instructions including trigger conditions, executors, and resource requirements.

4. The method for using the multi-scenario process-based emergency system based on the power emergency knowledge base according to claim 1, characterized in that: The knowledge graph construction in step (2) includes: integrating the emergency task instruction set using the GridGraph graph engine; integrating the power emergency professional dictionary through entity disambiguation and linking technology to generate a knowledge base containing more than 5 million entities and more than 1 million relationships.

5. The method for using a multi-scenario process-based emergency system based on a power emergency knowledge base according to claim 1, characterized in that: The dynamic generation of emergency response plans in step (3) includes: matching historical scenarios based on the KNN improved algorithm of the case library; updating decision content through real-time data feedback and dynamically adjusting emergency resource requirements.

6. The usage method of the multi-scenario process-based emergency system based on the power emergency knowledge base according to claim 1, characterized in that: The generation of the emergency response plan in step (4) includes: building a dynamic job responsibility database, integrating meteorological and power grid operation data; and using a deep belief network DBN and a collaborative filtering algorithm to achieve hybrid recommendation of instructions.

Images