Multi-modal large-scale corpus collection method in electricity handling service scene

Abstract

The invention discloses a multi-modal large-scale corpus collection method in an electricity handling service scene, which comprises the following steps: corpus data collection: corpus data related to an electricity handling service is collected from multiple sources, the corpus data sources cover all links of the electricity handling service, corpus data preprocessing is carried out, and corpus data collection is carried out; data standardization, data cleaning, data annotation and data storage actions, corpus model construction and multi-modal corpus model construction are performed on the collected corpus data, so that the corpus models are trained in various power utilization environments, and through fine tuning, the corpus models are adapted to different power utilization environments, power grid operation prediction results are obtained, power grid data are integrated, and power grid operation prediction is realized. And connecting the corpus model with a power grid real-time monitoring system to obtain an actual operation state result of the power grid.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal processing, corpus collection or corpus establishment, and more specifically, to a multimodal large-scale corpus collection method in an electricity service scenario. Background Art

[0002] In modern power systems, with the development of smart grids, power services are becoming more and more intelligent and personalized. In order to better serve users, a large amount of corpus data needs to be collected. These data may include but are not limited to customer service conversation records, equipment operation guides, user feedback, power grid operation data, etc. These data are not only diverse in form (text, voice, image, etc.), but also huge in quantity and widely distributed, involving all aspects of power services. Therefore, how to effectively collect, process and utilize these multimodal large-scale corpus data has become a key issue. Summary of the invention

[0003] In order to make full use of the corpus data in the power system, the present invention provides a multimodal large-scale corpus collection method in a power service scenario.

[0004] The technical solution of the present invention is as follows:

[0005] A multimodal large-scale corpus collection method in the electricity service scenario.

[0006] Corpus data collection: collect corpus data related to electricity service from multiple sources. The corpus data sources cover all aspects of electricity service.

[0007] Corpus data preprocessing, data standardization, data cleaning, data labeling and data storage of the collected corpus data,

[0008] Corpus model construction: construct a multimodal corpus model to train the corpus model in a variety of power consumption environments. Through fine-tuning, each corpus model can adapt to different power consumption environments and obtain grid operation prediction results.

[0009] Power grid data integration, connecting the corpus model with the real-time monitoring system of the power grid to obtain the actual operating status of the power grid,

[0010] Feedback, compare the grid operation prediction results with the actual grid operation status results, and apply the comparison results to the iterative training of the corpus model.

[0011] Collect corpus data from different channels, such as user consultation records, customer service conversations, power grid operation logs, equipment maintenance reports, etc. The corpus data covers a wide range, including but not limited to all power application, installation, maintenance, payment and other power service links. Standardize the data format of the collected corpus data to ensure that the corpus data from different sources can be compatible and interrelated, and remove erroneous or useless information to improve data quality. Centralize the scattered corpus data into a database (establish a corpus) for management and analysis. Then convert the format of the corpus data and annotate the corpus data, such as marking the fault type, user demand category and other information. After the above process is completed, use a suitable database management system to store the preprocessed corpus data. Construct a multimodal corpus model and use deep learning or other machine learning technologies to train the corpus model so that it can make accurate predictions under different power consumption environments. By fine-tuning the model parameters, the model can better adapt to the specific power consumption environment and provide accurate power grid operation prediction results. The above corpus model is integrated with the real-time monitoring system of the power grid, so that the model can receive real-time data from the power grid. The data stream of the real-time monitoring system of the power grid can include information such as current and voltage values, load conditions, and equipment working status. The differences between the results predicted by the model and the actual operating status of the power grid are compared, and these comparison results are used to adjust the model parameters, perform iterative training, and continuously optimize the model performance, so as to help the corpus model gradually improve its prediction ability, thereby more accurately predicting possible problems in the power grid.

[0012] In the above-mentioned large-scale corpus collection method under the electricity service scenario, in the corpus data collection stage,

[0013] Determine the source of the corpus data, determine the power service environment and power grid equipment where the data can be obtained, and form a corresponding list of corpus data sources.

[0014] Determine the corpus collection tools, and determine the corpus collection tools one by one according to the data transmission path of the power service, the connection status of the power grid equipment and the status of the data collection terminal, and check the corpus data source list.

[0015] Corpus data storage, setting up an initial corpus database, and transferring the corpus data collected or collected by the corpus collection tool to the initial corpus database for storage.

[0016] The above-mentioned large-scale corpus collection method in the electricity service scenario includes the following data standardization contents:

[0017] (1) Unified timestamp format;

[0018] (2) Unified unit format;

[0019] (3) Unified language expression;

[0020] (4) Unify data formats.

[0021] Furthermore, in the process of unifying the data format, the corpus data is cut and divided into multiple equal-length segments, or the keywords and related data in the corpus data are extracted according to the recorded conventional keyword library.

[0022] The above-mentioned large-scale corpus collection method in the electricity service scenario includes the following contents:

[0023] (1) Remove duplicates; (2) Fill in missing values; (3) Remove outliers.

[0024] The above-mentioned large-scale corpus collection method in the electricity service scenario includes the following data annotation contents:

[0025] Determine the annotation objects, including device status, customer service sentiment, and event type.

[0026] Formulate annotation rules, define annotation names and annotation mapping tables or text guidelines according to the annotation objects, and map the contents of the annotation objects to the annotation names according to the annotation mapping tables.

[0027] Determine the annotation tool, import the corpus data and standard rules into the annotation tool, and the annotation tool will create an annotation project and publish the annotation task.

[0028] Get the annotation results.

[0029] The above-mentioned large-scale corpus collection method in the electricity service scenario, the corpus model construction includes

[0030] Extract text features, extract annotated text segments from corpus data, convert data annotated text into word frequency vectors, and establish a mapping relationship between data annotated text and word frequency vectors.

[0031] Select a suitable model based on the existing model type and specific application environment, use network search or random determination to set model parameters, and establish a corpus model.

[0032] Import the word frequency vector into the corpus model,

[0033] The comparison results between the power grid operation prediction results and the actual power grid operation status results are applied to the iterative training of the corpus model to update the model parameters.

[0034] Furthermore, if the prediction result is a binary result, a logistic regression model or a vector machine model is selected as the corpus model.

[0035] Furthermore, if there are time series features in the corpus data, a recurrent neural network model is selected as the corpus model.

[0036] Furthermore, if the amount of corpus data to be processed is large, a convolutional neural network model or a Transformer architecture model is selected as the corpus model.

[0037] The present invention according to the above scheme has the beneficial effect that:

[0038] (1) The present invention ensures the consistency and comparability of the data by standardizing the collected corpus data, which is convenient for subsequent data analysis and model training. The data cleaning process removes noise and invalid information, improves the purity of the data, and enhances the accuracy of model training. Data annotation adds semantic information, which enables the machine learning model to better understand the data content and improves the model's interpretation ability. The data storage method after standardization, cleaning and annotation ensures the security and accessibility of the data, which is convenient for long-term maintenance and expansion.

[0039] (2) The present invention integrates the trained corpus model with the real-time monitoring system of the power grid, which can timely obtain the actual operating status of the power grid, enhance the timeliness and accuracy of the prediction results, and identify the deviation of the model by comparing the prediction results with the actual operating status, which can then be used for iterative training of the model to gradually improve the prediction accuracy. DETAILED DESCRIPTION

[0040] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0041] A multimodal large-scale corpus collection method in the electricity service scenario.

[0042] Corpus data collection: collect corpus data related to electricity service from multiple sources. The corpus data sources cover all aspects of electricity service.

[0043] Corpus data preprocessing, data standardization, data cleaning, data labeling and data storage of the collected corpus data,

[0044] Corpus model construction: construct a multimodal corpus model to train the corpus model in a variety of power consumption environments. Through fine-tuning, each corpus model can adapt to different power consumption environments and obtain grid operation prediction results.

[0045] Power grid data integration, connecting the corpus model with the real-time monitoring system of the power grid to obtain the actual operating status of the power grid,

[0046] Feedback, compare the grid operation prediction results with the actual grid operation status results, and apply the comparison results to the iterative training of the corpus model.

[0047] Collect corpus data from different channels, such as user consultation records, customer service conversations, power grid operation logs, equipment maintenance reports, etc. The corpus data covers a wide range, including but not limited to all power application, installation, maintenance, payment and other power service links. Standardize the data format of the collected corpus data to ensure that the corpus data from different sources can be compatible and interrelated, and remove erroneous or useless information to improve data quality. Centralize the scattered corpus data into a database (establish a corpus) for management and analysis. Then convert the format of the corpus data and annotate the corpus data, such as marking the fault type, user demand category and other information. After the above process is completed, use a suitable database management system to store the preprocessed corpus data. Construct a multimodal corpus model and use deep learning or other machine learning technologies to train the corpus model so that it can make accurate predictions under different power consumption environments. By fine-tuning the model parameters, the model can better adapt to the specific power consumption environment and provide accurate power grid operation prediction results. The above corpus model is integrated with the real-time monitoring system of the power grid, so that the model can receive real-time data from the power grid. The data stream of the real-time monitoring system of the power grid can include information such as current and voltage values, load conditions, and equipment working status. The differences between the results predicted by the model and the actual operating status of the power grid are compared, and these comparison results are used to adjust the model parameters, perform iterative training, and continuously optimize the model performance, so as to help the corpus model gradually improve its prediction ability, thereby more accurately predicting possible problems in the power grid.

[0048] In the corpus data collection phase,

[0049] Determine the source of the corpus data, determine the power service environment and power grid equipment where the data can be obtained, and form a corresponding list of corpus data sources.

[0050] Determine the corpus collection tools, and determine the corpus collection tools one by one according to the data transmission path of the power service, the connection status of the power grid equipment and the status of the data collection terminal, and check the corpus data source list.

[0051] Corpus data storage, setting up an initial corpus database, and transferring the corpus data collected or collected by the corpus collection tool to the initial corpus database for storage.

[0052] The sources of corpus data include:

[0053] (1) Customer service records such as customer service hotline records, chat records of online customer service platforms, electronic forms or emails submitted by customers, etc.

[0054] Customer service records will be saved. In the existing network environment, these customer service records will be uploaded and sent to the corresponding server and storage through the network. The data transmission path of these customer service records is an online path and can be easily obtained through network tools.

[0055] Usually, power companies will build customer service systems. This type of system will record the call records of all incoming customers. Use Python to write a script to call the API interface of the customer service center system to obtain the latest call records (even including call recording materials, etc.) regularly (for example, every hour) and save them to the local database. Similarly, for online customer service platforms, such as instant messaging tools on the company's official website, chat records can also be obtained through the API interface. For feedback information in emails, use a similar method to obtain and parse the email content regularly through the API interface provided by the mail server to extract useful information.

[0056] (2) Equipment operation records such as smart grid equipment (such as smart meters and sensors), SCADA systems, and equipment maintenance logs.

[0057] The operating status of power grid equipment is under monitoring in daily work, and the monitoring system of power grid equipment is connected to obtain equipment operation records.

[0058] The power company has deployed a SCADA system that collects data from all smart meters and sensors. Through the API interface of the SCADA system, the equipment operation status data can be obtained in real time. You can write a Python script to call the API interface of the SCADA system regularly (for example, every 5 minutes) to obtain the latest equipment operation status data and save it to the local database. In addition, equipment maintenance logs are usually saved in the company's internal database. These data can be obtained through the database query interface. Using the database query tool, log data can be regularly exported from the database and formatted for subsequent analysis.

[0059] (3) Market information such as power market trading platforms, government policy documents, and industry association reports.

[0060] Market information is open information and exists directly in public documents on the Internet and public platforms.

[0061] Use Python's requests library and BeautifulSoup library to write a web crawler script to regularly grab the latest market price information from the official website of the power market trading platform and save it to the local database. For policy documents issued by the government, web crawler technology is also used to regularly download the latest files from the official website, parse and save them.

[0062] (4) Weather information such as weather forecast data and historical weather records provided by the meteorological department.

[0063] Meteorological information contains undisclosed data and regular public information. Usually, power companies will cooperate with the local meteorological department. The meteorological department provides an API interface. A script is written using Python's requests library to call the API interface provided by the meteorological department regularly (for example, every hour) to obtain the latest weather forecast data and save it to the local database to obtain local weather forecast data in real time. You can also download weather records for a period of time from the meteorological department's historical database. You don't even need detailed meteorological information. It is also possible to obtain public meteorological information on the Internet through web crawler tools.

[0064] Data standardization includes

[0065] (1) Unified timestamp format; (2) Unified unit format; (3) Unified language representation; (4) Unified data format.

[0066] In the process of unifying the data format, the corpus data is cut and divided into multiple equal-length segments, or the keywords and related data in the corpus data are extracted according to the recorded regular keyword library.

[0067] According to the preset mapping table, the correspondence between the original corpus data representation words, the corpus field representation words and the update field representation words in the mapping table is obtained, and the cut or extracted corpus data is replaced with the update field representation words.

[0068] Unify the timestamp format so that the time of the collected corpus data is expressed in the same format, such as the time of the power status data collected by the equipment, the customer service registration time, the email time, etc. These times may use different time representation formats according to the requirements of different systems or devices, or there are differences in time zones, or there are hourly expressions, etc. Therefore, it is necessary to unify the time format in the corpus data and use the same expression method.

[0069] The unit formats of the physical and chemical parameters in the collected corpus data, such as voltage, resistance, temperature, length, etc., may use units of different systems. In order to facilitate and accurately calculate the subsequent simple numbers, the units of these parameters are unified.

[0070] The unification of language expression refers to the unification of computer language expression, such as some front-end data formats and protocol data formats. Due to the differences in data acquisition systems and web pages, these data formats are different, which will affect subsequent data analysis and calculations.

[0071] The unification of data formats is to facilitate calculations. Computer languages ​​are often expressed in English and other ways. Especially in program code calculations, the use of Chinese is minimized as much as possible in language representation, or Chinese is matched with English, numbers, and symbols in advance, and keywords in the corpus data are replaced (establishing a replacement language library to achieve replacement mapping relationships).

[0072] In one embodiment, equipment operation records from different sensors are collected from multiple sources, and the time format, unit and data format of each sensor record are inconsistent. There are also text information (customer service records) that record the interaction between customers and customer service. This information may contain text descriptions of the equipment status, as well as market information such as electricity market prices and demand forecasts from different databases. Among them, the log record of a certain device has a recorded time format of "2023-08-1512:00:00". Through the datetime library of the above Python script, we convert this timestamp into the ISO8601 format "2023-08-15T12:00:00Z"; the recorded power is 100 kilowatts (kW), which is converted to watts (W) through the conversion function, that is, 100000 watts. In a customer service record, there is a message "The equipment is running normally". We convert this message into a structured format, for example: {"status":"The equipment is running normally"}. ; The data in a certain market information is a numerical value stored in the form of a string, such as "100". We use `pd.to_numeric()` to convert it to a numerical type.

[0073] In the mapping table, the keywords (fields) of customer service records include customer_id, timestamp, message, etc., the keywords (fields) of equipment operation records include device_id, timestamp, voltage, current, etc., the keywords (fields) of market information include timestamp, price, demand, etc., and the keywords (fields) of weather information include timestamp, temperature, humidity, etc.

[0074] Through these steps, we can ensure that all collected data remains in a consistent format, which is convenient for subsequent analysis and processing without worrying about differences in data formats.

[0075] Data cleaning includes

[0076] (1) Remove duplicates; (2) Fill in missing values; (3) Remove outliers.

[0077] Deduplication is to remove duplicate records in a data set to ensure that each piece of data is unique, avoid misleading analysis, and reduce the burden on system operation. First, you need to determine which fields or columns are key fields to determine whether data records are duplicates. For example, in the operation records of power equipment, you can use timestamps, equipment IDs, etc. as key fields, and then use the pandas library in Python to remove duplicates.

[0078] Specific Python script code example:

[0079] import pandas as pd

[0080] # Load data

[0081] df=pd.read_csv('device_data.csv')

[0082] #Find duplicates

[0083] duplicates=df.duplicated(subset=['device_id','timestamp'],keep='first')

[0084] #Remove duplicates

[0085] df_unique = df[~duplicates]

[0086] #Save the processed data

[0087] df_unique.to_csv('device_data_cleaned.csv',index=False)

[0088] Filling missing values ​​is to deal with missing values ​​in the data set, ensure data integrity, and avoid bias in analysis. First, you need to identify which fields have missing values. You can use the isnull() method in pandas to find missing values. Then use mean / median filling (for numerical data, use the mean or median of the same type of data to fill), mode filling (for categorical data, you can use the mode to fill), forward / backward filling (for time series data, use forward filling or backward filling methods), interpolation (for time series data, use linear interpolation or other interpolation methods).

[0089] Specific Python script code example:

[0090] import pandas as pd

[0091] # Load data

[0092] df=pd.read_csv('device_data.csv')

[0093] #Find missing values

[0094] missing_values=df.isnull().sum()

[0095] #Fill numeric data with mean

[0096] df['voltage'].fillna(df['voltage'].mean(),inplace=True)

[0097] #Fill categorical data with majority

[0098] df['status'].fillna(df['status'].mode()[0],inplace=True)

[0099] #Fill time series data with forward filling

[0100] df['current'].fillna(method='ffill',inplace=True)

[0101] #Fill time series data with linear interpolation

[0102] df['temperature'].interpolate(inplace=True)

[0103] #Save the processed data

[0104] df.to_csv('device_data_filled.csv',index=False)

[0105] Deleting outliers is to identify and delete or correct outliers in the data set that do not conform to the norm, so as to avoid outliers affecting the analysis results. Specifically, statistical methods (such as IQR quartile range) or visualization methods (such as box plots) are used to identify outliers. For obviously abnormal data, they are directly deleted. For slightly abnormal data, they are corrected by methods such as replacing them with the mean or median, or limiting values ​​that exceed a certain range to a reasonable range.

[0106] Specific Python script code example:

[0107] import pandas as pd

[0108] import numpy as np

[0109] # Load data

[0110] df=pd.read_csv('device_data.csv')

[0111] # Calculate IQR

[0112] Q1=df['voltage'].quantile(0.25)

[0113] Q3=df['voltage'].quantile(0.75)

[0114] IQR=Q3-Q1

[0115] #Determine the range of outliers

[0116] lower_bound=Q1-1.5*IQR

[0117] upper_bound=Q3+1.5*IQR

[0118] # Remove outliers

[0119] df_cleaned=df[(df['voltage']>=lower_bound)&(df['voltage']<=

[0120] upper_bound)]

[0121] #Save the processed data

[0122] df_cleaned.to_csv('device_data_cleaned.csv',index=False)

[0123] Data annotation includes

[0124] Determine the annotation objects, including device status, customer service sentiment, and event type.

[0125] Formulate annotation rules, define annotation names and annotation mapping tables or text guidelines according to the annotation objects, and map the contents of the annotation objects to the annotation names according to the annotation mapping tables.

[0126] Determine the annotation tool, import the corpus data and standard rules into the annotation tool, and the annotation tool will create an annotation project and publish the annotation task.

[0127] Get the annotation results.

[0128] In the power environment, the objects of annotation may include but are not limited to annotation of equipment operating status (such as normal, warning or fault), customer service sentiment annotation (annotation of emotions in customer service records, such as positive, negative or neutral), and event annotation (annotation of the occurrence of specific events, such as equipment maintenance, power outages, etc.).

[0129] According to the above-mentioned annotation objects, define the relevant annotation names of the annotation objects, such as equipment status annotation (such as "normal", "warning", "fault"), customer service sentiment annotation (such as "positive", "negative", "neutral"), and event annotation (such as "equipment maintenance", "power outage accident", etc.). Then formulate a annotation mapping table or text guide. If the voltage and current in the corpus data are within the normal range, it is marked as "normal". If the voltage or current exceeds the normal range but does not reach the fault level, it is marked as "warning". If the voltage or current exceeds the fault threshold, it is marked as "fault". If there are positive words in the customer feedback of the corpus data, it is marked as "positive". If there are negative words in the customer feedback, it is marked as "negative". If there are no obvious positive or negative words in the customer feedback, it is marked as "neutral". If the corpus data mentions equipment maintenance, it is marked as "equipment maintenance". If the corpus data mentions power outage, it is marked as "power outage accident".

[0130] Select a data annotation tool suitable for the power environment. Commonly used annotation tools include ELAN, Label Studio, etc. ELAN is suitable for text and audio annotation, and Label Studio is suitable for various types of data annotation, supporting images, text, audio, etc. Import the corpus data and standard rules into the annotation tool. The annotation tool will create an annotation project and publish the annotation task. After the manual or artificial intelligence completes the annotation of the corpus data, the annotation results will be returned for review and confirmation.

[0131] Corpus model construction includes

[0132] Extract text features, extract annotated text segments from corpus data, convert data annotated text into word frequency vectors, and establish a mapping relationship between data annotated text and word frequency vectors.

[0133] Select a suitable model based on the existing model type and specific application environment, use network search or random determination to set model parameters, and establish a corpus model.

[0134] Import the word frequency vector into the corpus model,

[0135] The comparison results between the power grid operation prediction results and the actual power grid operation status results are applied to the iterative training of the corpus model to update the model parameters.

[0136] If the prediction result is a binary result, select the logistic regression model or vector machine model as the corpus model.

[0137] If there are time series features in the corpus data, select the recurrent neural network model as the corpus model.

[0138] If the amount of corpus data to be processed is large, choose a convolutional neural network model or a Transformer architecture model as the corpus model.

[0139] In a specific embodiment:

[0140] Example 1: Customer Service Satisfaction Survey Analysis Select a logistic regression model. The logistic regression model is simple, easy to understand and implement, and is suitable for handling binary or multi-classification problems. It can provide the degree of influence of each feature on the prediction result, which is very helpful for understanding which factors affect customer satisfaction. Moreover, when the amount of data is not particularly large, the logistic regression model can still give a good prediction result.

[0141] Example 2: Equipment Failure Prediction Select the vector machine model, which can handle high-dimensional data well and show good generalization ability in small sample conditions. Through the kernel trick, the vector machine can handle nonlinear classification problems and is suitable for complex situations such as equipment failure. In addition, the vector machine model has a certain tolerance for noise data and can better deal with outliers in equipment operation data.

[0142] Example 3: A recurrent neural network model is selected for customer service conversation analysis. The recurrent neural network model can process sequence data and is suitable for analyzing contextual information in customer service conversations. It also has memory capabilities and can remember previous inputs, which is very important in customer service conversations because conversations are sequential.

[0143] Example 4: Equipment Status Monitoring Report Analysis Select the convolutional neural network model, which has local perception characteristics and can capture local features in the text, which is very useful for the description in the equipment status monitoring report. In addition, the convolutional neural network model can automatically extract features, reduce the workload of manual feature engineering, perform parallel computing, and is suitable for processing large amounts of data.

[0144] Example 5: Market Information Analysis and Forecasting The Transformer architecture model is selected. The Transformer architecture model realizes parallel computing through the self-attention mechanism, which greatly speeds up the training speed. At the same time, the Transformer can capture long-distance dependencies in the text and can process complex language structures and semantic information, which is suitable for processing complex patterns in market information analysis.

[0145] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multimodal large-scale corpus collection method in an electricity service scenario, characterized in that: Corpus data collection: collect corpus data related to electricity services from multiple sources. The corpus data sources cover all aspects of electricity services. Corpus data preprocessing, data standardization, data cleaning, data labeling and data storage of the collected corpus data, Corpus model construction: construct a multimodal corpus model to train the corpus model in a variety of power consumption environments. Through fine-tuning, each corpus model can adapt to different power consumption environments and obtain grid operation prediction results. Power grid data integration, connecting the corpus model with the real-time monitoring system of the power grid to obtain the actual operating status of the power grid, Feedback, compare the grid operation prediction results with the actual grid operation status results, and apply the comparison results to the iterative training of the corpus model.

2. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 1, it is characterized in that: In the corpus data collection phase, Determine the source of the corpus data, determine the power service environment and power grid equipment where the data can be obtained, and form a corresponding list of corpus data sources. Determine the corpus collection tools, and determine the corpus collection tools one by one according to the data transmission path of the power service, the connection status of the power grid equipment and the status of the data collection terminal, and check the corpus data source list. Corpus data storage, setting up an initial corpus database, and transferring the corpus data collected or collected by the corpus collection tool to the initial corpus database for storage.

3. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 1, it is characterized in that: Data standardization includes (1) Unified timestamp format; (2) Unify the unit format; (3) Unify language expressions; (4) Unify data formats.

4. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 3, it is characterized in that: In the process of unifying the data format, the corpus data is cut and divided into multiple equal-length segments, or the keywords and related data in the corpus data are extracted according to the recorded regular keyword library. According to the preset mapping table, the correspondence between the original corpus data representation words, the corpus field representation words and the update field representation words in the mapping table is obtained, and the cut or extracted corpus data is replaced with the update field representation words.

5. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 1, it is characterized in that: Data cleaning includes (1) Remove duplicates; (2) Fill in missing values; (3) Remove outliers.

6. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 1, it is characterized in that: Data annotation includes Determine the annotation objects, including device status, customer service sentiment, and event type. Formulate annotation rules, define annotation names and annotation mapping tables or text guidelines according to the annotation objects, and map the contents of the annotation objects to the annotation names according to the annotation mapping tables. Determine the annotation tool, import the corpus data and standard rules into the annotation tool, and the annotation tool will create an annotation project and publish the annotation task. Get the annotation results.

7. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 1, it is characterized in that: Corpus model construction includes Extract text features, extract annotated text segments from corpus data, convert data annotated text into word frequency vectors, and establish a mapping relationship between data annotated text and word frequency vectors. Select a suitable model based on the existing model type and specific application environment, use network search or random determination to set model parameters, and establish a corpus model. Import the word frequency vector into the corpus model, The comparison results between the power grid operation prediction results and the actual power grid operation status results are applied to the iterative training of the corpus model to update the model parameters.

8. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 7, it is characterized in that: If the prediction result is a binary result, select the logistic regression model or vector machine model as the corpus model.

9. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 7, it is characterized in that: If there are time series features in the corpus data, select the recurrent neural network model as the corpus model.

10. According to the method for collecting multimodal large-scale corpus in the electricity service scenario described in claim 7, it is characterized in that: If the amount of corpus data to be processed is large, choose a convolutional neural network model or a Transformer architecture model as the corpus model.