Information extraction method and system for power demand prediction

Abstract

The invention belongs to the field of electric power, and discloses an information extraction method and system for electric power demand prediction, and the method comprises the steps: collecting multi-modal original data related to the electric power demand prediction from a plurality of data sources, and carrying out the cleaning, screening and structural processing of the data, so as to form a preliminary screening sample set; based on the modal type of the preliminary screening sample set and the power business scene, generating Rubrics assessment rules including a plurality of quantifiable dimensions; performing multi-dimensional quality scoring on the data in the preliminary screening sample set according to the Rubrics evaluation rules, and performing classification screening on the data according to a scoring result and a threshold value; and adjusting the weight, the scoring tendency and the screening threshold value of the Rubrics evaluation detailed rule based on the precision feedback of the prediction model and the manual labeling result so as to realize closed-loop optimization of information extraction.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of electric power, and particularly relates to an information extraction method and system for electric power demand prediction. BACKGROUND

[0002] Electric power demand prediction is the core basis of power grid dispatching optimization and resource allocation, and its accuracy directly affects the safety and economy of power grid operation. With the continuous increase of new energy installed capacity ratio, the electric power system is facing the challenges of high volatility, diversified load and complex market mechanism. Precise prediction needs to rely on the effective input of multi-dimensional information, including policy documents, meteorological data, industrial dynamics and other multi-modal data.

[0003] However, the traditional information extraction method has weak multi-modal adaptation ability, can only process text or structured data, cannot effectively fuse multi-modal information such as images and time series, and has dynamic iteration lag, relies on artificial preset rules, and is difficult to adapt to dynamic needs such as changes in new energy policy and industrial scene. In addition, the traditional information extraction method lacks a quantitative evaluation system, resulting in low information screening efficiency and strong subjectivity. SUMMARY

[0004] Therefore, the present application discloses an information extraction method and system for electric power demand prediction, which can solve the problems in the related art.

[0005] To achieve the above-mentioned purpose, the technical solutions of the present application are as follows: According to the first aspect of the present application, an information extraction method for electric power demand prediction is provided, which comprises: Collecting multi-modal raw data related to electric power demand prediction from multiple data sources, and cleaning, screening and structuring the data to form a preliminary screening sample set; Based on the modal type and electric power business scene of the preliminary screening sample set, generating Rubrics evaluation rules containing multiple quantifiable dimensions; According to the Rubrics evaluation rules, multi-dimensional quality scores are given to the data in the preliminary screening sample set, and the data is classified and screened according to the score results and threshold values; Based on the accuracy feedback of the prediction model and the artificial annotation results, the weight, scoring tendency and screening threshold of the Rubrics evaluation rules are adjusted to realize closed-loop optimization of information extraction.

[0006] According to the second aspect of the present application, an information extraction device for electric power demand prediction is provided, which comprises: The acquisition module is used for collecting multi-modal raw data from multiple data sources and performing cleaning, screening and structuring; A generating module is configured to dynamically generate Rubrics evaluation rules specific to the power scenario based on the processed data; A scoring module is configured to quantitatively score and classify the data according to the rules; An optimization module is configured to dynamically adjust system parameters according to feedback to realize closed-loop learning.

[0007] According to a third aspect of the present application, an electronic device is provided, comprising: a processor; a memory for storing processor-executable instructions; The processor implements the steps of the method according to the first aspect by running the executable instructions.

[0008] According to a fourth aspect of the present application, a computer-readable storage medium is provided, which stores computer instructions that, when executed by a processor, implement the steps of the method according to the first aspect.

[0009] As can be seen from the above technical solutions, the information extraction method for power demand prediction disclosed in the present application has the following advantages: On the one hand, high-value information is filtered through quantitative evaluation rules to establish an explicit correlation between information quality and prediction accuracy, thereby improving the quality of prediction input. On the other hand, the closed-loop optimization mechanism adapts to changes in power grid business, enhances the dynamic adaptive ability of the system, and reduces the maintenance cost of artificial rules. In addition, the present application supports unified processing of text, image, and time series data, thereby reducing the workload of manual screening. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 is a framework diagram of an information extraction system for power demand prediction provided by an exemplary embodiment; Figure 2 is a flowchart of an information extraction method for power demand prediction provided by an exemplary embodiment; Figure 3 is a schematic structural diagram of a device provided by an exemplary embodiment; Figure 4 is a block diagram of an information extraction device for power demand prediction provided by an exemplary embodiment. DETAILED DESCRIPTION

[0011] The exemplary embodiments will be described in detail herein with reference to the attached drawings. The description below refers to the accompanying drawings, which show, by way of example, specific embodiments with which one or more embodiments of the present application can be implemented. The following description, however, is not representative of all possible embodiments consistent with one or more embodiments of the present application. Rather, it is merely an example of apparatus and methods consistent with some aspects of one or more embodiments of the present application as detailed in the appended claims.

[0012] It is noted that the steps of the methods of the respective embodiments are not necessarily performed in the order as shown and described herein. In some other embodiments, the methods can comprise more or less steps than described herein. Also, a single step described herein can be broken down into multiple steps in other embodiments, and multiple steps described herein can be combined into a single step in other embodiments.

[0013] The accuracy of power demand forecasting essentially depends on the effective input of multi-dimensional power-related information. From policy documents (electricity price adjustment, energy consumption control), meteorological data (high temperature, strong wind), to industrial dynamics (plant expansion, shutdown), and power grid structured data, the completeness and accuracy of this information directly determines the effectiveness of the prediction model. Power information extraction for power demand forecasting is a key link to bridge the gap between "data" and "accurate prediction". Only by efficiently extracting high-value information from massive multi-modal data can we provide reliable support for prediction models, and the technical level directly affects the upper limit of the accuracy of power demand forecasting. Its core feature is "scene-specific": unlike general information extraction, it needs to deeply adapt to the knowledge of the power industry and the logic of demand forecasting to ensure that the extracted information can directly serve the prediction business.

[0014] The current information extraction method applied in the power industry is mostly based on general models or traditional rules, and there are significant pain points in adapting to the power demand forecasting scenario. First, the multi-modal adaptation capability is weak. Traditional methods are mostly aimed at single modalities such as text policies, and cannot effectively adapt to multi-modal data such as images and time series, resulting in "missing useful information and misextracting irrelevant information". Second, the dynamic adaptability is insufficient. The factors affecting power grid consumption have dynamic characteristics (such as the introduction of emerging new energy policies and the rise of new industrial scenarios), and the traditional manual preset rule method cannot be iterated in time, resulting in a continuous decline in extraction effectiveness as the business changes. Third, the quality evaluation system is missing. There is currently a lack of information quality quantitative evaluation system for the power scenario, making it difficult to automatically determine the completeness, timeliness, and relevance of the extracted information. In actual application, it needs to rely on manual screening, which is not only inefficient, but also easy to miss high-value information due to subjective judgment.

[0015] As the core basis of power grid dispatching optimization and resource allocation, the accuracy of power demand forecasting directly affects the safety, economy and stability of power grid operation. In recent years, the Rubrics-based quality assessment and feedback learning method has rapidly developed in the field of intelligent content generation and complex decision evaluation, and has become an important technical means to improve the reliability of model output due to its multi-dimensional, quantifiable and iterative characteristics. However, this method has not been widely applied in power grid information extraction and forecasting tasks. Power grid data is highly dynamic and scenario-specific, and the Rubrics framework is theoretically suitable for establishing an explicit link between extraction quality and forecasting accuracy. However, there is a lack of adaptive design for power business, which is a technical gap worth exploring in the industry.

[0016] In related technologies, there are still significant deficiencies in power information extraction methods: first, most algorithms are based on general models or artificial preset rules, which lack specific adaptation to power scenarios and are difficult to cover policy texts, weather time series, industrial operation data and other multi-source information; second, power consumption factors change rapidly with policy and industry, while traditional rule systems update slowly, leading to invalid extraction results; third, there is a lack of systematic quality quantification evaluation mechanism, relying on manual screening and subjective judgment, which is low in efficiency and prone to missing valuable information. The above problems restrict the widespread application of information extraction in power demand forecasting, especially in the context of the continuous increase in new energy installed capacity and the volatile power spot market, the lack of extraction quality has become a key bottleneck to improve forecasting accuracy.

[0017] To solve the dual problems of information extraction adaptability and evaluation in related technologies, and fill the technical gap of Rubrics in power grid scenarios, the present application proposes an information extraction method and system for power demand forecasting.

[0018] Rubrics is a structured scoring tool that breaks down evaluation objectives into observable and quantifiable multi-dimensional standards, and sets different performance levels and specific descriptions for each dimension, used for objective, consistent and transparent quality judgment of evaluation objects. This information extraction framework is a multi-source information extraction solution that adapts to the characteristics of the power industry. It first classifies multi-modal data such as policy documents, industrial images and weather time series according to power grid scenarios, then uses small network automatic learning to learn the correlation between data and power demand forecasting, generates customized rubrics evaluation rules, and finally combines LLM to complete information quality quantification scoring and screening, which can efficiently and accurately extract valuable core information for power demand forecasting from massive heterogeneous data. Therefore, the Rubrics-based power information extraction framework has the following values in power demand forecasting: Establish a direct correlation between information quality and prediction accuracy: The framework quantitatively evaluates the effectiveness, relevance, and quantifiability of extracted information through the rubrics system, clearly defines the influence weight of information quality in different dimensions on the prediction result, and provides core support for subsequent optimization of prediction model input and establishment of "information quality-prediction accuracy" mapping relationship.

[0019] Adapt to the dynamic development needs of power grid business: The rubrics evaluation rules do not require manual presetting and support dynamic iteration. Through the closed-loop mechanism of "prediction accuracy feedback-weight optimization-model update", the evaluation dimensions of emerging scenarios such as new energy policy and green hydrogen can be quickly incorporated, solving the problem of traditional manual rule iteration lag, and ensuring that information extraction always adapts to the dynamic changes of power grid business under the background of energy transformation.

[0020] Realize accurate selection and efficient processing of multi-modal information, greatly reduce manual dependence: The framework first adapts text, image, time series and other multi-modal data through power scene-specific classification, and then clearly defines the quality standards of each type of information with customized rubrics, and completes automatic quantitative scoring and screening with LLM. It not only solves the problem of "single modality incomplete coverage, useful information missing / irrelevant information missing" in traditional methods, but also replaces the inefficient mode of manual review one by one, which can reduce more than 80% of the workload of manual screening, and adapt to the rapid processing needs of massive heterogeneous data in power grid.

[0021] Figure 1 is a framework diagram of an information extraction system for power demand forecasting provided by an example embodiment. As Figure 1 shown, the information extraction system for power demand forecasting includes: a data collection and management module (i.e., a collection module) for collecting multi-modal raw data from multiple data sources and performing cleaning, screening, and structured processing; a rubrics rule generation module (i.e., a generation module) for dynamically generating Rubrics evaluation rules specific to power scenarios based on the processed data; an information quality scoring and screening module (i.e., a scoring module) for quantitatively scoring and classifying and screening data according to the rules; a dynamic evaluation and iterative optimization module (i.e., an optimization module) for dynamically adjusting system parameters according to feedback to realize closed-loop learning.

[0022] Figure 2 is a flowchart of an information extraction method for power demand forecasting provided by an example embodiment. As Figure 2 shown, the method can be applied to the control unit of the information extraction system for power demand forecasting as Figure 2 shown; can include the following steps: Step 201, collecting multi-modal raw data related to power demand prediction from multiple data sources, and cleaning, screening and structuring the data to form a preliminary screening sample set; Step 202, generating Rubrics evaluation rules containing multiple quantifiable dimensions based on the modal type of the preliminary screening sample set and the power business scenario; Step 203, performing multi-dimensional quality scoring on the data in the preliminary screening sample set according to the Rubrics evaluation rules, and classifying and screening the data according to the scoring results and threshold values; Step 204, adjusting the weight, scoring tendency and screening threshold of the Rubrics evaluation rules based on the precision feedback of the prediction model and the artificial annotation results to realize closed-loop optimization of information extraction.

[0023] In this embodiment, on the one hand, high-value information is screened through quantitative evaluation rules, an explicit correlation between "information quality-prediction accuracy" is established, thereby improving the prediction input quality; on the other hand, through the closed-loop optimization mechanism, the system dynamic adaptive ability is enhanced, and the artificial rule maintenance cost is reduced; in addition, the application supports unified processing of text, image and time series data, and reduces the artificial screening workload. An end-to-end automatic information extraction and optimization process is provided, which systematically integrates multi-modal data processing, dynamic rule generation, quantitative scoring screening and closed-loop optimization, fundamentally improves the quality, reliability and timeliness of power demand prediction input information, and realizes the transformation from "experience screening" to "quantitative evaluation and dynamic optimization".

[0024] In an embodiment, the multi-modal raw data related to power demand prediction is collected from multiple data sources, and the data is cleaned, screened and structured, including: obtaining raw information from power grid internal structured data, external environmental data, unstructured document and image data sources; performing preliminary static screening based on task keyword matching on the collected data to eliminate information irrelevant to the prediction target; performing data cleaning on the screened data, including missing value filling, noise elimination and duplicate value deduplication; extracting key fields and adding modal labels, and storing all data in a unified schema format.

[0025] The data collection and management module is used to organize raw data from multiple sources and convert it into a set of structured data that has been cleaned of noise, standardized in format, and only retains information that is valuable for the prediction task, providing high-quality input for downstream processing modules. The raw data includes internal structured data of the power grid (load curves, electricity consumption, industry statistics, etc.), external environmental data (meteorological platform observation data, policy documents, news reports, economic indicators, etc.), unstructured data (PDF files, social media posts, etc.), and multi-modal data (numerical tables, text, images, etc.). Such data has problems such as inconsistent format, noise interference, or missing fields due to the diversity of sources.

[0026] The data collection and management process is divided into three steps: (1) preliminary static screening: based on task keyword matching and domain standard rules, the collected multi-source data is filtered in the first round, and irrelevant information to the prediction target is removed; (2) data cleaning: through methods such as missing value filling (such as interpolation or mean substitution), noise removal (such as 3σ method or anomaly detection model), and duplicate value deduplication, the data quality is guaranteed; (3) data structuring: key fields such as source institution, publication date, and impact field are extracted, and modal labels and task labels are added, and all data are stored in a unified schema format (such as JSON document or database table), ensuring that subsequent modules can be directly called.

[0027] The final output is a set of high-quality data sets that can be directly input into downstream scene classification or prediction models, each record contains a unique ID, source, time, modal type, and task label, and its quality and relevance have been guaranteed in the first round, preventing irrelevant or low-quality data from entering the subsequent processing process.

[0028] In this embodiment, through the standardized data preprocessing process, the quality and consistency of multi-source heterogeneous data are ensured, providing high-quality, standardized input for subsequent analysis, and reducing noise and irrelevant information to the prediction model from the source.

[0029] In an embodiment, the Rubrics evaluation rules containing a plurality of quantifiable dimensions are generated, including: generating the Rubrics evaluation rules based on a pre-constructed rule generation model; wherein the dimensions of the Rubrics evaluation rules include general quality dimensions and power grid business specific dimensions; the general quality dimensions include structural integrity, timeliness, domain relevance, accuracy, and verifiability; the power grid business specific dimensions set specific indicators for text, time series / numerical, and image modal data.

[0030] The Rubrics generation module is used to generate scoring rules for the multi-source and multi-modal preliminary screening samples processed by the data collection and management module to adapt to the power demand prediction scenario. The core function of the module is to dynamically construct scoring dimensions and rules from the keyword library of power demand influencing factors and business templates according to the data modal type (structured power grid data, external environment data, unstructured documents, images, etc.), source field (internal power grid, external power environment), and user-defined evaluation requirements.

[0031] In the rule construction process, the dimension setting refers to the actual needs of the power grid business, including but not limited to: structural integrity: whether the information field or data point is complete; timeliness: the matching degree of the information timestamp and the prediction task window (recent weather, latest policy); field relevance: the business relevance to the power demand prediction target; accuracy: the correctness of the content and value; verifiability: reliable and traceable source.

[0032] And add special dimensions for different modalities: text: keyword coverage, logical consistency, topic matching degree; time series / numerical: sampling integrity, abnormal point proportion, statistical feature rationality; image: resolution and recognizability, key meta-information integrity.

[0033] To train high-quality rule generation models, the module uses the SFT (Supervised Fine-Tuning) method, which uses artificially designed high-quality Rubrics or high-level model-generated Rubrics rules that have been manually screened as training samples. The model input is scenario description, modal information, and user requirements, and the output is structured Rubrics (JSON or table format). The fine-tuning process enables the model to learn to generate accurate, comprehensive, and interpretable scoring rules under different data modalities and business scenarios, ensuring long-term consistency between the rules and actual power grid needs.

[0034] In this embodiment, the pre-trained model dynamically generates evaluation rules that combine generality and power business characteristics, enabling quality evaluation standards to accurately adapt to the characteristics of different modal data and the specific needs of power prediction scenarios, improving the relevance and accuracy of evaluation.

[0035] In one embodiment, the multi-dimensional quality scoring of the data in the preliminary screening sample set according to the Rubrics evaluation rules, and classifying and filtering the data according to the scoring results and threshold values, includes: calling a large language model as a scoring model, and performing dimension-by-dimension evaluation according to the Rubrics evaluation rules to output dimension-by-dimension scores; weighting the dimension-by-dimension scores to obtain a comprehensive score; comparing the comprehensive score with a pre-set scoring threshold, and dividing the data into a high-quality data set, a to-be-reviewed data set, and a rejected data set according to the comparison result.

[0036] Further, the score threshold is dynamically adjusted by a double-objective optimization algorithm of acceptance rate and manual review pass rate.

[0037] The module is used for quality scoring and screening of the preliminary screening dataset to generate a high-quality dataset that can be directly used for prediction tasks. The module receives the scoring rubrics output by the Rubrics guideline generation module and scores each piece of data in multiple dimensions according to the rubrics, covering general indicators such as structural integrity, timeliness, field relevance, accuracy, and verifiability, as well as modality-specific indicators. The scores of each dimension are weighted and aggregated to calculate a comprehensive score. The module sets threshold standards based on the scoring results: samples with a comprehensive score higher than the set threshold directly enter the power grid operation analysis and dispatch tool, samples with a score in the threshold interval enter the list to be reviewed, and samples with a score lower than the threshold are rejected or archived for backup. The threshold can be dynamically adjusted according to historical data distribution and prediction accuracy feedback to adapt to the quality requirements of different business scenarios. The output of the module includes a high-quality sample dataset, a dataset to be reviewed, and rejected sample records, ensuring the reliability and controllability of prediction input.

[0038] In this embodiment, on the one hand, the large language model is used for automatic and multi-dimensional fine scoring, and through weighted aggregation and threshold determination, objective quantification and efficient hierarchical screening of information quality are realized, reducing the workload and subjectivity of manual screening; on the other hand, by dynamically adjusting the scoring threshold, the system can intelligently balance the "quality" and "quantity" of information screening, adapt to changes in the demand for information quality and quantity at different business stages, and enhance the flexibility and practicality of the system.

[0039] In an embodiment, the closed loop includes three-way feedback optimization: weight optimization, model optimization, and threshold optimization; wherein the weight optimization updates the weight matrix of the Rubrics rubrics according to the correlation between the dimension score and the prediction accuracy; the model optimization corrects the output distribution of the scoring model by statistics of the misjudgment proportion in different intervals; and the threshold optimization dynamically adjusts the threshold standard according to the business load and the acceptance rate.

[0040] The dynamic evaluation and iterative optimization module is used for self-adaptive adjustment of Rubrics weight, scoring model tendency and scoring threshold according to the performance of the information scoring system in long-term operation and the artificial marking results. The module collects the scoring output and artificial confirmation data of each batch of tasks, evaluates the contribution of each Rubrics dimension through correlation analysis and regression calculation, outputs a new weight matrix and feeds back to the criterion generation model; at the same time, the false rejection rate of the scoring model in different score intervals is counted to correct the deviation of the output distribution; at the system level, a new scoring threshold is generated through a double-target optimization algorithm of acceptance rate and false rejection rate, balancing the number and quality of input samples. All the updated results are reviewed and confirmed by professionals and take effect in the next round of tasks. The module enables the system to have continuous learning and adaptive ability, ensuring that the scoring rules, model tendency and threshold criteria remain consistent with business needs at different stages.

[0041] In this embodiment, a multi-dimensional closed-loop optimization mechanism is constructed, which can simultaneously adjust the evaluation criteria, scoring model and screening threshold, ensuring that the system can continuously improve itself and maintain high evaluation performance in the long term.

[0042] In an embodiment, the method further comprises: inputting the high-quality information set output by the scored screening step into a power demand prediction model to improve the prediction accuracy of power grid dispatching and resource allocation.

[0043] In this embodiment, the specific application value of the high-quality information set is clearly defined, directly linking the information extraction link with the final power demand prediction task, forming a complete value closed loop from "information optimization" to "prediction accuracy improvement".

[0044] The present application quantifies indicators such as information integrity, timeliness and field relevance into computable dimensions by introducing the Rubrics evaluation system, establishes an explicit correlation between information quality and prediction accuracy, realizes the transition of prediction input from "experience selection" to "quantifiable optimization", and improves the accuracy and reliability of prediction model input data.

[0045] The dynamic evaluation and iterative optimization module automatically updates the criterion weight, scoring model tendency and scoring threshold by analyzing the correlation between Rubrics dimension scores and prediction results, realizes the closed-loop iteration of "prediction feedback-weight optimization-re-evaluation", and enables the system to learn and correct itself as the business environment and energy policy change, maintaining long-term stable evaluation performance.

[0046] The system realizes the automatic quantitative evaluation and screening of information quality under the Rubrics criterion framework, which can reduce a large amount of manual review work. Compared with the traditional extraction method relying on manual judgment, the present application can reduce more than 80% of the manual screening workload, and improve the processing speed and operation and maintenance efficiency of massive power heterogeneous data.

[0047] Since the scoring dimensions, weight coefficients, and thresholds are all explicitly configurable parameters, the application provides clear traceable evaluation basis and result interpretation paths, facilitating experts to verify the model decision-making process. Meanwhile, the Rubrics criteria and dynamic optimization mechanism can flexibly add evaluation dimensions and business templates, supporting the rapid access and expansion of emerging scenarios (such as green electricity hydrogen production and demand response markets).

[0048] Figure 3 is a schematic structural diagram of a device provided by an example embodiment. Please refer to Figure 3 At the hardware level, the device includes a processor 302, an internal bus 304, a network interface 306, a memory 308, and a non-volatile memory 310, and of course, other hardware required for functions. One or more embodiments of the application can be implemented in a software manner, such as reading a corresponding computer program from the non-volatile memory 310 into the memory 308 by the processor 302 and then running. Of course, in addition to the software implementation, one or more embodiments of the application do not exclude other implementation manners, such as logic devices or a combination of software and hardware, etc. That is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or a logic device.

[0049] Please refer to Figure 4 An information extraction device for power demand prediction can be applied to a device as shown in Figure 4 to implement the technical solutions of the application, and the device includes: An acquisition unit 401 is configured to acquire multi-modal original data related to power demand prediction from multiple data sources, and perform cleaning, screening, and structured processing on the data to form a preliminary screening sample set; A generation unit 402 is configured to generate a Rubrics evaluation rule containing multiple quantifiable dimensions based on the modal types of the preliminary screening sample set and power business scenarios; A scoring unit 403 is configured to perform multi-dimensional quality scoring on the data in the preliminary screening sample set according to the Rubrics evaluation rule, and classify and screen the data according to the scoring results and thresholds; An adjustment unit 404 is configured to adjust the weights, scoring tendencies, and screening thresholds of the Rubrics evaluation rule based on the accuracy feedback of the prediction model and the artificial annotation results, to realize closed-loop optimization of information extraction.

[0050] Optionally, the acquisition unit 401 is specifically configured to: obtain original information from power grid internal structured data, external environmental data, unstructured document, and image data sources; Performing preliminary static screening based on task keyword matching on the collected data to eliminate irrelevant information from the predicted target; Performing data cleaning on the screened data, including missing value filling, noise removal, and duplicate value removal; Extracting key fields and adding modal labels, and structuring all data in a unified schema format.

[0051] Optionally, the production unit 402 is specifically configured to: Generate the Rubrics evaluation rules based on the pre-constructed rule generation model; wherein the dimensions of the Rubrics evaluation rules include general quality dimensions and grid business-specific dimensions; the general quality dimensions include structural integrity, timeliness, field relevance, accuracy, and verifiability; the grid business-specific dimensions set specific indicators for text, time series / numerical, and image modal data.

[0052] Optionally, the scoring unit 403 is specifically configured to: Call a large language model as a scoring model, and perform dimension-by-dimension evaluation according to the Rubrics evaluation rules to output dimension scores; Weighted calculation of the dimension scores to obtain a comprehensive score; Compare the comprehensive score with a pre-set score threshold, and divide the data into a high-quality data set, a to-be-reviewed data set, and a rejected data set according to the comparison result.

[0053] Further, the score threshold is dynamically adjusted by a double-objective optimization algorithm of acceptance rate and manual review pass rate.

[0054] Optionally, the closed loop includes three-way feedback optimization: weight optimization, model optimization, and threshold optimization; wherein the weight optimization updates the weight matrix of the Rubrics rules according to the correlation between the dimension scores and the prediction accuracy; the model optimization corrects the output distribution of the scoring model according to the misjudgment proportion in different intervals; and the threshold optimization dynamically adjusts the threshold standard according to the business load and the acceptance rate.

[0055] Optionally, the device further comprises: The input unit 405 is configured to input the high-quality information set output by the scoring and screening step into the power demand prediction model to improve the prediction accuracy of grid dispatching and resource allocation.

[0056] The systems, apparatuses, modules, or units illustrated in the above embodiments can be specifically implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer, and the specific form of the computer can be a personal computer, a laptop computer, a cellular phone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0057] In a typical configuration, a computer includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0058] The memory can include non-persistent memory in computer readable media, random access memory (RAM), and / or non-volatile memory such as read-only memory (ROM) or flash memory (flash RAM). The memory is an example of computer readable media.

[0059] The computer readable media includes permanent and non-permanent, removable and non-removable media, which can be implemented by any method or technology to store information. The information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read only memory (ROM), electrically erasable programmable read only memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, disk storage, quantum memory, graphene-based storage medium or other magnetic storage device, or any other non-transmission medium that can be used to store information accessible by a computing device. According to the definition in this paper, computer readable media does not include transitory computer readable media, such as modulated data signals and carriers.

[0060] For the computer readable medium (or computer readable storage medium) as described above or any other form, computer instructions can be stored thereon, which are executed by a processor to implement one or more of the above embodiments, thereby realizing the technical solutions of the present application.

[0061] The present application also provides a computer program, which is executed by a processor to implement one or more of the above embodiments, thereby realizing the technical solutions of the present application. The computer program can be specifically recorded on the computer readable medium as described above or any other form, and the present application does not limit this.

[0062] It is also important to note that the terms "comprises", "comprising", or other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus. An element proceeded by "comprises... a" does not, without more constraints, exclude the presence of additional identical elements in the process, method, article, or apparatus that comprises the element.

[0063] The above described embodiments of the application have been described in connection with what are presently considered to be the most practical and preferred embodiments from a present practical viewpoint. However, it is to be understood that the application is not to be limited to the disclosed embodiments, but instead, is intended to cover various arrangements of the present application that are different from those described, which fall within the scope of the appended claims and their equivalents.

[0064] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of one or more embodiments of the application. As used herein, the singular forms "a", "an" and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms "comprises" and / or "comprising," when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0065] It is to be understood that the singular forms "a", "an", and "the" include plural referents unless the context clearly dictates otherwise. Preferred embodiments of the application are described herein, including the best mode known to the inventors of practicing the application. It is to be understood that the application is not to be limited to the specific embodiments disclosed and that modifications will occur to those skilled in the art to which the application pertains without departing from the spirit of the application. Accordingly, the application is not to be limited by the specific examples presented herein, but is to be given the full scope of the appended claims and equivalents thereof.

[0066] The foregoing is considered as illustrative only of the principles of the application. Further, since numerous modifications and changes will readily occur to those skilled in the art, it is not desired to limit the application to the exact construction and operation described. Accordingly, all such variations are intended to be included within the scope of the present application as defined in the claims below and their equivalents.

Claims

1. An information extraction method for electricity demand forecasting, characterized in that, The method includes: Multimodal raw data related to electricity demand forecasting are collected from multiple data sources, and the data is cleaned, filtered, and structured to form an initial screening sample set. Based on the modal types and power business scenarios of the initial screening sample set, Rubrics evaluation rules containing multiple quantifiable dimensions are generated. The data in the initial screening sample set are scored in multiple dimensions according to the Rubrics evaluation criteria, and the data are classified and screened according to the scoring results and thresholds. The weights, scoring biases, and screening thresholds of the Rubrics evaluation criteria are adjusted based on the accuracy feedback of the prediction model and the results of manual annotation, so as to achieve closed-loop optimization of information extraction.

2. The method according to claim 1, characterized in that, The process of collecting multimodal raw data related to electricity demand forecasting from multiple data sources, and cleaning, filtering, and structuring the data, includes: Raw information is obtained from structured data within the power grid, external environmental data, unstructured documents, and image data sources. The collected data is subjected to preliminary static screening based on task keyword matching to remove information that is irrelevant to the prediction target; Data cleaning is performed on the filtered data, including missing value imputation, noise removal, and duplicate value removal; Extract key fields and add modal labels, then store all data in a structured manner according to a unified schema format.

3. The method according to claim 1, characterized in that, The generation of Rubrics evaluation rules, which includes multiple quantifiable dimensions, includes: The Rubrics evaluation rules are generated based on a pre-built rule generation model. The Rubrics evaluation rules include general quality dimensions and power grid business-specific dimensions. The general quality dimensions include structural integrity, timeliness, domain relevance, accuracy, and verifiability. The power grid business-specific dimensions set specific indicators for different modal data such as text, time series / numerical, and image.

4. The method according to claim 1, characterized in that, The process of performing multi-dimensional quality scoring on the data in the initial screening sample set according to the Rubrics evaluation criteria, and classifying and filtering the data based on the scoring results and thresholds, includes: The large language model is used as the scoring model, and the Rubrics evaluation rules are used to evaluate each dimension to output the score for each dimension. The scores of each dimension are weighted to obtain a comprehensive score; The comprehensive score is compared with the pre-set score threshold, and the data is divided into high-quality datasets, datasets to be reviewed, and datasets to be removed based on the comparison results.

5. The method according to claim 4, characterized in that, The scoring threshold is dynamically adjusted using a dual-objective optimization algorithm that considers both acceptance rate and manual review pass rate.

6. The method according to claim 1, characterized in that, The closed loop includes three-way feedback optimization: weight optimization, model optimization, and threshold optimization. Specifically, the weight optimization updates the weight matrix of the Rubric rules based on the correlation between dimensional scores and prediction accuracy. The model optimization calculates the misjudgment ratio of the scoring model in different intervals and corrects the deviation of its output distribution. The threshold optimization dynamically adjusts the threshold standard based on business load and acceptance rate.

7. The method according to claim 1, characterized in that, The method further includes: The high-quality information set output from the scoring and screening process is input into the power demand forecasting model to improve the forecasting accuracy of grid dispatching and resource allocation.

8. An information extraction system for electricity demand forecasting, characterized in that, The system includes: The data acquisition module is used to collect multimodal raw data from multiple data sources and perform cleaning, filtering, and structuring processing. The generation module is used to dynamically generate Rubric evaluation rules specific to power scenarios based on the processed data; The scoring module is used to quantitatively score and classify the data according to the aforementioned rules. The optimization module is used to dynamically adjust system parameters based on feedback to achieve closed-loop learning.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor implements the steps of the method as described in any one of claims 1-7 by running the executable instructions.

10. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, this instruction implements the steps of the method as described in any one of claims 1-7.

Images