Electric power abnormal data risk grading and handling method and system

Abstract

The invention belongs to the technical field of artificial intelligence, and particularly relates to a power abnormal data risk grading and handling method and system, and the method comprises the following specific steps: collecting data from a multi-source channel, and carrying out the preprocessing and feature engineering of the collected data; building a classifier based on a plurality of decision tree models, and selecting information gains as evaluation indexes to construct decision trees; finely tuning the locally deployed GLM open source large model; disposing suggested retrieval enhanced questions and answers; according to the method, various information sources and the processing knowledge base are combined in real time, more comprehensive and accurate processing suggestions are provided, and real-time updating of the processing knowledge base and dynamic generation of the processing suggestions are achieved through a retrieval enhancement scheme.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method and system for grading and handling power abnormal data risks. Background Art

[0002] The current power system is facing a complex and ever-changing operating environment. With the rapid development of smart grid technology, the power system is generating a massive amount of operating data, including various key parameters such as voltage, current, and power. These data cover all aspects of the power system operation and provide a comprehensive information basis for monitoring and analyzing the power grid status.

[0003] However, the operating environment of the power system is becoming increasingly complex. On the one hand, power demand and load are constantly fluctuating due to social and economic development and changes in user behavior. On the other hand, the large-scale grid connection of renewable energy has also brought more uncertainties to the operation of the power grid. These changes have made the operating status of the power system more complex and changeable.

[0004] In this case, it is crucial to identify abnormal conditions in power data in a timely manner. Once equipment failure or abnormal operation occurs, if it is not discovered and countermeasures are not taken in time, it may cause serious power grid accidents and threaten the safe and stable operation of the power grid. Therefore, how to quickly and accurately detect abnormal conditions from massive power operation data is a key technical issue to ensure the safety and reliability of the power system. There are some problems with the existing risk identification and disposal methods. First, traditional risk identification relies too much on manual experience judgment and is difficult to deal with massive and complex power data. The power system involves a wide variety of data, including various key parameters such as voltage, current, power, frequency, etc. The amount of data is huge, and traditional manual analysis methods are difficult to fully grasp. Secondly, most of the existing risk disposal measures are passive, with slow response speed, making it difficult to achieve real-time and efficient risk control. Furthermore, the existing risk classification standards are often too rough and cannot provide specific response guidance for risks of different types and severity. These problems restrict the further improvement of the risk management and control capabilities of the power system.

[0005] In order to solve the above problems, a method and system for grading and handling the risk of abnormal power data is proposed. The core of the invention is to orchestrate the workflow and collaboration strategies between multiple AI models to build an intelligent risk assessment and response mechanism. Specifically, the system integrates multiple AI models such as anomaly detection, risk assessment, and disposal strategy generation, and realizes collaborative analysis through workflow orchestration. This multi-AI model collaborative approach can give full play to the advantages of each model, improve the accuracy and efficiency of risk identification and response, and provide processing suggestions.

[0006] In addition, the present invention has intelligent and automated response capabilities. It can analyze power data in real time, automatically identify abnormal situations, evaluate risk levels, and give corresponding disposal suggestions to achieve fast and efficient risk response. At the same time, the system provides detailed classification standards and targeted disposal suggestions for abnormal risks of different types and severity, which greatly improves the meticulousness and comprehensiveness of risk classification. At the same time, by timely updating the disposal suggestion knowledge base, the real-time nature of the large model disposal suggestions can be continuously updated to achieve continuous improvement of knowledge-driven. Summary of the invention

[0007] In view of the above and / or existing problems in a method and system for grading and handling power abnormality data risks, the present invention is proposed.

[0008] Therefore, the purpose of the present invention is to provide a method and system for grading and handling power abnormality data risks, which can solve the above-mentioned existing problems.

[0009] To solve the above technical problems, according to one aspect of the present invention, the present invention provides the following technical solutions:

[0010] A method for grading and handling power abnormal data risks includes the following specific steps:

[0011] S1: Collect data from multiple sources and perform preprocessing and feature engineering on the collected data;

[0012] S2: Build a classifier based on multiple decision tree models and select information gain as the evaluation indicator to construct the decision tree;

[0013] S3: Fine-tune the local deployment of the GLM open source large model;

[0014] S4: Disposition suggestion retrieval enhanced question answering;

[0015] S5: Process control.

[0016] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, the specific steps of preprocessing and feature engineering in S1 are as follows:

[0017] Step 1: Normalize the numerical features to obtain normalized data;

[0018] Step 2: Record the time series features, and construct feature values ​​by sliding average of the recorded time series features. The average of multiple values ​​is used to construct feature values ​​to reduce errors and data volume.

[0019] Step 3: Segment the Chinese data, and then use TF-IDF technology to calculate the importance of the word in the document and use it as a feature;

[0020] Step 4: Convert categorical features to integers for model processing;

[0021] Step 5: Manually label samples and label different data features with corresponding risk levels;

[0022] Step 6: Divide the processed features and risk level labels into training sets and test sets for model training and evaluation;

[0023] Step 7: Save the data to the database and set the initial abnormal threshold for the data.

[0024] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, the decision tree in S2 is a tree structure composed of nodes and directed edges, each internal node represents a test on an attribute, each branch represents a test result, each leaf node represents a category or output value, and the path from the root node to the leaf node corresponds to a decision rule.

[0025] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, each decision tree is constructed by selecting information gain as an evaluation index, and the calculation formula of its information entropy is as follows:

[0026]

[0027] Among them, Y represents the classification level; p k represents the proportion of the kth class of samples in the current sample set D; assuming that the discrete attribute a has V possible values ​​{a 1 ,a 2 ,…,a V}, if a is used to divide the sample set D, V branch nodes will be generated; among them, the vth branch node contains all the samples in D whose value on attribute a is a v The sample is recorded as D v ; At this time, D can be calculated according to the information entropy calculation formula v The information entropy of the branch nodes is given weights considering that different branch nodes contain different numbers of samples. That is, the branch nodes with more samples have a greater impact, so the information gain obtained by dividing the sample set D with attribute a can be calculated:

[0028]

[0029] Since the larger the information gain, the more clustered the categories obtained by dividing with attribute a will be, therefore, information gain can be used to select the division attributes of the decision tree, and the formula for selecting the attributes is:

[0030] a * = argmax a∈A Gain(D,a)

[0031] Select the attribute that can maximize the information gain, recursively select the optimal feature, and split the training data according to its feature. If these subsets can be basically correctly classified, then build leaf nodes and divide their subsets into corresponding leaf nodes; if there are still subsets that cannot be basically correctly classified, then select new optimal features for their subsets, continue to split them, and build corresponding nodes. Continue recursively in this way until all training data subsets are basically correctly classified.

[0032] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, the specific steps of S3 are as follows:

[0033] S31: Download the open source big model to the local computer, load the local big model, build the abnormal data, risk classification and corresponding cause analysis data set, and ensure that each abnormal data has a clear cause label;

[0034] S32: Use lightweight fine-tuning technology to fine-tune the open source large model so that it has the ability to analyze and answer the causes of abnormal data.

[0035] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, the specific steps of S4 are as follows:

[0036] S41: Write a document on risk level and abnormal cause handling suggestions based on expert advice and historical documents;

[0037] S42: Load the processing suggestion file, parse the document content, use the Jieba word segmentation method to segment the document collection, obtain the word list of each document, and build an inverted list for each word, create a dictionary, the key is the word, the value is the document ID containing the word and its location information, then score the query and document, and obtain the corresponding page number and content of the document with the highest query score;

[0038] S43: When providing suggestions, the corresponding disposal plan file is retrieved based on the reasons answered by the S3 model, and the corresponding disposal suggestion content is obtained to construct prompt input. The large model provides disposal suggestions based on the retrieved disposal suggestion content.

[0039] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, wherein: S42 is to score the document and the query using the BM25 algorithm, wherein the BM25 algorithm is a widely used retrieval model for estimating the relevance of the document to a given query;

[0040] The formula of the BM25 algorithm is:

[0041]

[0042] Q is the query, d is the document, q i is the term in the query, f(q i , d) is the frequency of the term in document d, |d| is the length of document d, avgdl is the average document length of the document collection, k1 and b are adjustable parameters, usually the value of k1 is between 1.2 and 2, the value of b is between 0.5 and 0.8, IDF(q i ) is the term q i The inverse document frequency is usually calculated as Where N is the total number of documents, n(q i ) is the term containing the term q i The number of documents.

[0043] As a preferred solution of the method for grading and handling power abnormal data risks described in the present invention, the specific steps of S5 are as follows:

[0044] S51: The data monitor monitors in real time whether the data exceeds the threshold range, and if it exceeds the threshold range, the risk level assessment model is activated to perform data assessment;

[0045] S52: Use the local large model of S3 to analyze the cause of the evaluated data and results, retrieve the processing suggestions based on the cause, and obtain preliminary document suggestions;

[0046] S53: Use preliminary suggestions as enhanced input to the larger model to give more complete answers.

[0047] A power abnormal data risk classification and disposal system, comprising:

[0048] Data collection and processing module, used to collect data from multiple source channels and perform preprocessing and feature engineering on the collected data;

[0049] Building module for building classifiers based on multiple decision tree models;

[0050] Fine-tuning module, used to fine-tune the local deployment of the GLM open source large model;

[0051] The processing module is used to process suggestion retrieval and enhance question answering;

[0052] The process control module is used to determine whether the data exceeds the threshold range and perform subsequent processing on the data if it exceeds the threshold range;

[0053] The data collection and processing module is connected to the construction module, the construction module is connected to the fine-tuning module, the fine-tuning module is connected to the disposal module, and the disposal module is connected to the process control module.

[0054] Compared with existing technologies:

[0055] The present invention realizes the provision of anomaly detection, risk grading, cause analysis and processing suggestions for real-time data by integrating learning decision tree models and dialogue generation of large models. In the data collection and processing links, the real-time and integrity of the data are ensured. The update capability of multiple decision tree models is integrated, and the decision tree models can be updated one by one in a timely manner after accumulating new data without affecting the overall operation of the system, thereby improving the accuracy and robustness of anomaly detection and risk grading. In terms of risk grading and cause analysis, integrated learning improves the generalization ability of the model and reduces the risk of overfitting by combining multiple classifiers. The decision tree model has strong interpretability due to its feature-based conditional judgment, which is convenient for users to understand the judgment basis of the model. The large model can better adapt to local data and specific application scenarios through fine-tuning, and improve the risk cause inspection and attribution capabilities of power abnormal data. Finally, a variety of information sources and processing knowledge bases are combined in real time to provide more comprehensive and accurate processing suggestions, and the real-time update of the processing knowledge base and the dynamic generation of disposal suggestions are realized through the retrieval enhancement scheme. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of the process of the present invention;

[0057] Figure 2 Schematic diagram of a single decision tree model example of the present invention. DETAILED DESCRIPTION

[0058] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0059] The present invention provides a method for grading and handling power abnormal data risks. Figure 1-Figure 2 , including the following specific steps:

[0060] S1: Collect data from multiple sources, such as voltage load, power consumption data, and temperature in transformers; or equipment behavior data in transformers, such as switch operation records and maintenance records; or environmental factors around transformers and other equipment, such as air humidity, wind speed, whether there is a thunderstorm, etc., and preprocess and feature engineer the collected data;

[0061] The specific steps of preprocessing and feature engineering in S1 are as follows:

[0062] Step 1: Normalize the numerical features to obtain normalized data;

[0063] Step 2: Record time series features, such as power consumption data, voltage load, and temperature, and construct feature values ​​by sliding average of the recorded time series features. The average of multiple values ​​is used to construct feature values ​​to reduce errors and data volume.

[0064] Step 3: Segment the Chinese data, and then use TF-IDF technology to calculate the importance of the word in the document and use it as a feature;

[0065] Step 4: Convert categorical features to integers for model processing;

[0066] Step 5: Manually label samples and label different data features with corresponding risk levels;

[0067] Step 6: Divide the processed features and risk level labels into training sets and test sets for model training and evaluation;

[0068] Step 7: Save the data to the database and set the initial abnormal threshold for the data. For example, if the voltage load data reaches x watts, the abnormal risk classification is started.

[0069] S2: Build a classifier based on multiple decision tree models and select information gain as the evaluation indicator to build a decision tree. The decision tree model is a widely used supervised learning algorithm that can be used for both classification and regression problems. Its core idea is to segment the data space through a series of rules to make predictions on unknown data.

[0070] The decision tree in S2 is a tree structure consisting of nodes and directed edges. Each internal node represents a test on an attribute, each branch represents a test result, each leaf node represents a category or output value, and the path from the root node to the leaf node corresponds to a decision rule, such as Figure 2 As shown;

[0071] Each decision tree is constructed by selecting information gain as the evaluation index. The calculation formula of its information entropy is as follows:

[0072]

[0073] Among them, Y represents the classification level. For example, if there are three risk levels, high risk, medium risk, and low risk, then Y is equal to 3; p k represents the proportion of the kth class of samples in the current sample set D; assuming that the discrete attribute a has V possible values ​​{a 1 ,a 2 ,…,a V}, if a is used to divide the sample set D, V branch nodes will be generated; among them, the vth branch node contains all the samples in D whose value on attribute a is a v The sample is recorded as D v ; At this time, D can be calculated according to the information entropy calculation formula v The information entropy of the branch nodes is given weights considering that different branch nodes contain different numbers of samples. That is, the branch nodes with more samples have a greater impact, so the information gain obtained by dividing the sample set D with attribute a can be calculated:

[0074]

[0075] Since the larger the information gain, the more clustered the categories obtained by dividing with attribute a will be, therefore, information gain can be used to select the division attributes of the decision tree, and the formula for selecting the attributes is:

[0076] a * = argmax a∈A Gain(D,a)

[0077] Select the attribute that can maximize the information gain, recursively select the optimal feature, and segment the training data according to its feature. If these subsets can be basically correctly classified, then construct leaf nodes and divide their subsets into corresponding leaf nodes; if there are still subsets that cannot be basically correctly classified, then select new optimal features for their subsets, continue to segment them, and construct corresponding nodes. This process is repeated recursively until all training data subsets are basically correctly classified.

[0078] Assume that there are 100 samples in data set D, of which 70 samples are "low risk", 20 samples are "medium risk", and 10 samples are "high risk";

[0079] Information entropy of D Ent(D):

[0080]

[0081] Ent(D)≈1.1568

[0082] Assume that the data set is divided into three categories according to the three values ​​of "voltage load": high: 20 samples, of which 5 are low risk, 10 are medium risk, and 5 are high risk; medium: 50 samples, of which 42 are low risk, 5 are medium risk, and 3 are high risk; low: 30 samples, of which 23 are low risk, 5 are medium risk, and 2 are high risk;

[0083] The information entropy Ent(D_high) of the sub-dataset D_high is:

[0084]

[0085] Ent(D h igh)=1.5

[0086] The information entropy Ent(D_medium) of the sub-dataset D_medium is:

[0087]

[0088] Ent(D m edium)≈0.787

[0089] The information entropy Ent(D_low) of the sub-dataset D_low is:

[0090]

[0091] Ent(D m edium)≈0.9848

[0092] Calculate the information gain for "voltage load":

[0093]

[0094] Gain(D,voltageload)≈0.1679

[0095] The information gain of the "voltage load" attribute is about 0.1679; the gains of other attributes are calculated using a similar method, and finally the attribute with the highest gain is selected as the basis for division;

[0096] Data training and division are performed on each decision tree to form multiple fully trained decision tree models. In the testing phase, the prediction results of multiple models are voted on and the result with the most votes is selected as the final prediction. At the same time, the feature importance of each decision tree model is analyzed, and the importance score of each feature is summarized and sorted to obtain an integrated feature importance score. The top n features are selected for cause analysis of the large model. With this integrated method, new data can be gradually accumulated and each decision tree model can be updated separately without affecting the overall risk assessment and provision of disposal recommendations.

[0097] S3: Fine-tune the local deployment of the GLM open source large model;

[0098] The specific steps of S3 are as follows:

[0099] S31: Download the open source big model (such as ChatGLM-6B) to the local computer, load the local big model, build the abnormal data, risk classification and corresponding cause analysis data set, and ensure that each abnormal data has a clear cause label, such as {"input": "Abnormal data description and risk classification 1", "output": "Corresponding cause 1"};

[0100] S32: Use lightweight fine-tuning technology to fine-tune the open source large model, such as using Lora technology to fine-tune the open source large model to enable it to have the ability to analyze and answer the causes of abnormal data;

[0101] S4: Disposition suggestion retrieval enhanced question answering;

[0102] The specific steps of S4 are as follows:

[0103] S41: Write a document on risk level and abnormal cause handling suggestions based on expert advice and historical documents;

[0104] S42: Load the processing suggestion file, parse the document content, use the Jieba word segmentation method to segment the document collection, obtain the word list of each document, and build an inverted list for each word, create a dictionary, the key is the word, the value is the document ID containing the word and its location information, then score the query and document, and obtain the corresponding page number and content of the document with the highest query score;

[0105] In S42, documents and queries are scored using the BM25 algorithm, which is a widely used retrieval model for estimating the relevance of documents to a given query.

[0106] The formula of the BM25 algorithm is:

[0107]

[0108] Q is the query, d is the document, q i is the term in the query, f(q i , d) is the frequency of the term in document d, |d| is the length of document d, avgdl is the average document length of the document collection, k1 and b are adjustable parameters, usually the value of k1 is between 1.2 and 2, the value of b is between 0.5 and 0.8, IDF(q i ) is the term q i The inverse document frequency is usually calculated as Where N is the total number of documents, n(q i ) is the term containing the term q i The number of documents;

[0109] S43: When providing suggestions, the corresponding disposal plan file is retrieved according to the reasons answered by the S3 model, and the corresponding disposal suggestion content is obtained to construct prompt input. The large model provides disposal suggestions based on the retrieved disposal suggestion content.

[0110] Assuming that you are an expert in power risk management, please combine the given information and give your suggestions for management. Please answer truthfully. If the suggestion cannot be found in the information, please answer "I don't know";

[0111] material:

[0112] -{information1}

[0113] -{information2}

[0114] -{information3}

[0115] Output the large model and retrieve the most similar text material link as the answer to the user to the question-answering robot interactive interface;

[0116] S5: Process control;

[0117] The specific steps of S5 are as follows:

[0118] S51: The data monitor monitors in real time whether the data exceeds the threshold range, and if it exceeds the threshold range, the risk level assessment model is activated to perform data assessment;

[0119] S52: Use the local large model of S3 to analyze the cause of the evaluated data and results, retrieve the processing suggestions based on the cause, and obtain preliminary document suggestions;

[0120] S53: Use preliminary suggestions as enhanced input to the larger model to give more complete answers.

[0121] A power abnormal data risk classification and disposal system, comprising:

[0122] Data collection and processing module, used to collect data from multiple source channels and perform preprocessing and feature engineering on the collected data;

[0123] Building module for building classifiers based on multiple decision tree models;

[0124] Fine-tuning module, used to fine-tune the local deployment of the GLM open source large model;

[0125] The processing module is used to process suggestion retrieval and enhance question answering;

[0126] The process control module is used to determine whether the data exceeds the threshold range and perform subsequent processing on the data if it exceeds the threshold range;

[0127] The data collection and processing module is connected to the construction module, the construction module is connected to the fine-tuning module, the fine-tuning module is connected to the disposal module, and the disposal module is connected to the process control module.

[0128] Although the present invention has been described above with reference to the embodiments, various modifications may be made thereto and parts thereof may be replaced by equivalents without departing from the scope of the present invention. In particular, as long as there is no structural conflict, the various features in the embodiments disclosed in the present invention may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for grading and handling power abnormal data risks, characterized in that: The specific steps are as follows: S1: Collect data from multiple sources and perform preprocessing and feature engineering on the collected data; S2: Build a classifier based on multiple decision tree models and select information gain as the evaluation indicator to construct the decision tree; S3: Fine-tune the local deployment of the GLM open source large model; S4: Disposition suggestion retrieval enhanced question answering; S5: Process control.

2. A method for grading and handling power abnormal data risks according to claim 1, characterized in that: The specific steps of preprocessing and feature engineering in S1 are as follows: Step 1: Normalize the numerical features to obtain normalized data; Step 2: Record the time series features, and construct feature values ​​by sliding average of the recorded time series features. The average of multiple values ​​is used to construct feature values ​​to reduce errors and data volume. Step 3: Segment the Chinese data, and then use TF-IDF technology to calculate the importance of the word in the document and use it as a feature; Step 4: Convert categorical features to integers for model processing; Step 5: Manually label samples and label different data features with corresponding risk levels; Step 6: Divide the processed features and risk level labels into training sets and test sets for model training and evaluation; Step 7: Save the data to the database and set the initial abnormal threshold for the data.

3. A method for grading and handling power abnormal data risks according to claim 1, characterized in that: The decision tree in S2 is a tree structure consisting of nodes and directed edges, each internal node represents a test on an attribute, each branch represents a test result, each leaf node represents a category or output value, and the path from the root node to the leaf node corresponds to a decision rule.

4. A method for grading and handling power abnormality data risks according to claim 3, characterized in that: Each decision tree is constructed by selecting information gain as an evaluation index, and the calculation formula of its information entropy is as follows: Among them, Y represents the classification level; p k represents the proportion of the kth class of samples in the current sample set D; assuming that the discrete attribute a has V possible values ​​{a 1 ,a 2 ,…,a V }, if a is used to divide the sample set D, V branch nodes will be generated; among them, the vth branch node contains all the samples in D whose value on attribute a is a v The sample is recorded as D v ; At this time, D can be calculated according to the information entropy calculation formula v The information entropy of the branch nodes is given weights considering that different branch nodes contain different numbers of samples. That is, the branch nodes with more samples have a greater impact, so the information gain obtained by dividing the sample set D with attribute a can be calculated: Since the larger the information gain, the more clustered the categories obtained by dividing with attribute a will be, therefore, information gain can be used to select the division attributes of the decision tree, and the formula for selecting the attributes is: a * =argmax a∈A Gain(D,a) Select the attribute that can maximize the information gain, recursively select the optimal feature, and split the training data according to its feature. If these subsets can be basically correctly classified, then build leaf nodes and divide their subsets into corresponding leaf nodes; if there are still subsets that cannot be basically correctly classified, then select new optimal features for their subsets, continue to split them, and build corresponding nodes. Continue recursively in this way until all training data subsets are basically correctly classified.

5. A method for grading and handling power abnormality data risks according to claim 1, characterized in that: The specific steps of S3 are as follows: S31: Download the open source big model to the local computer, load the local big model, build the abnormal data, risk classification and corresponding cause analysis data set, and ensure that each abnormal data has a clear cause label; S32: Use lightweight fine-tuning technology to fine-tune the open source large model so that it has the ability to analyze and answer the causes of abnormal data.

6. A method for grading and handling power abnormal data risks according to claim 1, characterized in that: The specific steps of S4 are as follows: S41: Write a document on risk level and abnormal cause handling suggestions based on expert advice and historical documents; S42: Load the processing suggestion file, parse the document content, use the Jieba word segmentation method to segment the document collection, obtain the word list of each document, and build an inverted list for each word, create a dictionary, the key is the word, the value is the document ID containing the word and its location information, then score the query and document, and obtain the corresponding page number and content of the document with the highest query score; S43: When providing suggestions, the corresponding disposal plan file is retrieved based on the reasons answered by the S3 model, and the corresponding disposal suggestion content is obtained to construct prompt input. The large model provides disposal suggestions based on the retrieved disposal suggestion content.

7. A method for grading and handling power abnormality data risks according to claim 6, characterized in that: In S42, the document and the query are scored by the BM25 algorithm, where the BM25 algorithm is a widely used retrieval model for estimating the relevance of the document to a given query; The formula of the BM25 algorithm is: Q is the query, d is the document, q i is the term in the query, f(q i , d) is the frequency of the term in document d, |d| is the length of document d, avgdl is the average document length of the document collection, k1 and b are adjustable parameters, usually the value of k1 is between 1.2 and 2, the value of b is between 0.5 and 0.8, IDF(q i ) is the term q i The inverse document frequency is usually calculated as Where N is the total number of documents, n(q i ) is the term containing the term q i The number of documents.

8. A method for grading and handling power abnormal data risks according to claim 1, characterized in that: The specific steps of S5 are as follows: S51: The data monitor monitors in real time whether the data exceeds the threshold range, and if it exceeds the threshold range, the risk level assessment model is activated to perform data assessment; S52: Use the local large model of S3 to analyze the cause of the evaluated data and results, retrieve the processing suggestions based on the cause, and obtain preliminary document suggestions; S53: Use preliminary suggestions as enhanced input to the larger model to give more complete answers.

9. A power abnormal data risk classification and handling system, characterized in that: include: Data collection and processing module, used to collect data from multiple source channels and perform preprocessing and feature engineering on the collected data; Building module for building classifiers based on multiple decision tree models; Fine-tuning module, used to fine-tune the local deployment of the GLM open source large model; The processing module is used to process suggestion retrieval and enhance question answering; The process control module is used to determine whether the data exceeds the threshold range and perform subsequent processing on the data if it exceeds the threshold range; The data collection and processing module is connected to the construction module, the construction module is connected to the fine-tuning module, the fine-tuning module is connected to the disposal module, and the disposal module is connected to the process control module.

Images