Meteorological report automatic generation system and method based on large language model

Abstract

This application provides an automated meteorological report generation system and method based on a large language model. The system includes: a data preprocessing module for acquiring raw meteorological numerical forecast data and performing structured transformation on the raw meteorological numerical forecast data to generate standardized meteorological data in a preset format; a knowledge base construction module for constructing a multi-source meteorological professional domain knowledge base; a prompt word construction module for receiving the standardized meteorological data and the multi-source meteorological professional domain knowledge base, configuring them in prompt words at different levels, and generating a unified model input instruction after concatenation; a parallel generation module for generating multiple candidate meteorological reports by using the model input instruction as a basis and calling the large language model in parallel through differentiated generation parameters; and a preferred output module for automatically evaluating the quality of the multiple candidate meteorological reports and outputting the final meteorological report based on the evaluation results.

Description

Technical Field

[0001] This invention relates to the fields of weather forecasting and artificial intelligence technology, and in particular to an automated weather report generation system and method based on a large language model. Background Technology

[0002] Numerical Weather Prediction (NWP) is a core technology in modern weather forecasting. Its basic principle is to treat the atmosphere as a fluid and, by solving the dynamic and thermodynamic equations describing atmospheric motion, extrapolate the atmospheric state over a future period using numerical integration on a computer, taking the current atmospheric state as the initial condition. The output of numerical prediction models is multi-dimensional gridded data, providing forecast values ​​for meteorological elements such as temperature, wind speed, air pressure, precipitation, and irradiance over time on a regular grid of points globally or regionally. This type of data is large in volume, high in dimensionality, and highly specialized, making it difficult for ordinary users to directly understand and use.

[0003] Large Language Models (LLMs) represent a significant breakthrough in artificial intelligence in recent years. They are deep learning models trained on massive amounts of text data, possessing powerful natural language understanding and generation capabilities. Representative products include the GPT series and DeepSeek. LLMs can generate fluent and coherent natural language text based on input text commands (i.e., prompts), and have been widely applied in scenarios such as text summarization, content generation, and question answering systems.

[0004] Weather forecast reports have significant applications across multiple industries. For example, in the new energy sector, wind and solar power generation are highly dependent on weather conditions; wind speed determines turbine output, and solar irradiance determines solar power generation. New energy operators need to regularly obtain professional weather forecast reports for power generation prediction and grid dispatch. In the aviation sector, aircraft takeoffs and landings and route planning heavily rely on accurate forecasts of meteorological elements such as wind fields, visibility, and thunderstorms. Aviation meteorological services need to regularly produce and publish airport and route forecast reports. All of these industries face a demand for large-scale, high-frequency weather report production, which traditional manual writing methods cannot meet.

[0005] Currently, the production process for weather forecast reports typically involves meteorological service agencies first obtaining gridded forecast data for the target area from numerical weather prediction models, then manually interpreting the data using their professional knowledge, and finally writing a forecast report in written form. After review, the report is distributed to clients. This process is highly dependent on professional personnel, has a long production cycle for a single report, and is limited by manpower, making it difficult to simultaneously support the production needs of multiple stations, multiple industries, and high-frequency reports.

[0006] Since the rise of large language models, research and engineering practices have attempted to use them for the automatic generation of meteorological texts. The basic idea is to take numerical forecast data or its summary as input, and guide the large language model to output the forecast report text through prompt words. For example, some systems directly concatenate the numerical values ​​of forecast elements into a text description and input it into the model, which then generates a complete report paragraph.

[0007] Current solutions employ a method that directly calls the large language model API, takes numerical weather prediction data summaries as input, and generates weather forecast text. The typical implementation steps are as follows: The first step is to obtain forecast data for the target area from numerical weather prediction models, and then manually or through a simple program to extract several key values ​​(such as maximum temperature, maximum wind speed, etc.). The second step is to concatenate the extracted numerical values ​​into a simple text description, which is then used as a prompt word input into the large language model. The third step is to call the large language model API to obtain the report text generated in a single instance, which is then used as the final output.

[0008] The above scheme has the following obvious shortcomings: First, the processing of numerical forecast data is crude, extracting only a small amount of data without systematic quality control and structure transformation, resulting in serious information loss; Second, the design of prompt words is simple and does not incorporate meteorological professional knowledge constraints, resulting in insufficient professionalism of the generated text and easy for forecast conclusions to be inconsistent with the numerical data; Third, it only performs a single generation without any quality assessment or optimization process, and the randomness of the large language model directly determines the quality of the final output, making stability impossible to guarantee and failing to meet the requirements of operational applications.

[0009] In summary, the generation of existing weather forecast reports, which relies on professional forecasters manually interpreting numerical model outputs and writing texts, has the following main drawbacks: 1. The raw grid data output by numerical weather prediction models is a multi-dimensional physical field, which lacks a standardized structured processing procedure. This results in forecasters having to spend a lot of time on manual interpretation, leading to low report production efficiency and an inability to meet the needs of multi-site, high-frequency business services. 2. When existing large language models are directly applied to generate meteorological reports, the lack of effective constraints from meteorological professional knowledge leads to insufficient professionalism in the generated content, and the forecast conclusions are prone to inconsistencies with numerical forecast data. 3. The generation process of large language models is random, and there is no quality control mechanism for each generation, which leads to unstable output report quality and makes it difficult to meet the reliability requirements of business applications. Summary of the Invention

[0010] The purpose of this invention is to provide an automated meteorological report generation system and method based on a large language model, aiming to solve the above-mentioned problems in the prior art.

[0011] This invention provides an automated meteorological report generation system based on a large language model, comprising: The data preprocessing module is used to acquire raw meteorological numerical forecast data and perform structured transformation on the raw meteorological numerical forecast data to generate standardized meteorological data in a preset format. The knowledge base construction module is used to build a multi-source meteorological professional knowledge base; The prompt word construction module, connected to the data preprocessing module and the knowledge base construction module, is used to receive the standardized meteorological data and the multi-source meteorological professional domain knowledge base, and configure them in prompt words at different levels, which are then concatenated to generate a unified model input instruction; The parallel generation module, connected to the prompt word construction module, is used to generate multiple candidate weather reports by calling the large language model in parallel through differentiated generation parameters based on the model input instructions. The preferred output module, connected to the parallel generation module, is used to perform automated quality assessment on the multiple candidate meteorological reports and output the final meteorological report based on the assessment results.

[0012] This invention provides a method for automatically generating meteorological reports based on a large language model, comprising: The raw meteorological numerical forecast data is obtained through the data preprocessing module, and the raw meteorological numerical forecast data is structurally transformed to generate standardized meteorological data in a preset format. A multi-source meteorological professional knowledge base is constructed using a knowledge base construction module; The standardized meteorological data and the multi-source meteorological professional domain knowledge base are received by the prompt word construction module and configured into prompt words at different levels. After being concatenated, a unified model input command is generated. Based on the input instructions of the model, the parallel generation module calls the large language model in parallel through differentiated generation parameters to generate multiple candidate weather reports. The multiple candidate meteorological reports are automatically evaluated using the optimized output module, and the final meteorological report is output based on the evaluation results.

[0013] This invention also provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it implements the steps of the above-described method for automatically generating meteorological reports based on a large language model.

[0014] This invention also provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, implements the steps of the above-described method for automatically generating meteorological reports based on a large language model.

[0015] The following beneficial effects can be achieved by adopting the embodiments of the present invention: The embodiments of the present invention propose an automated meteorological report generation system based on numerical model forecasting and multi-candidate selection of large language models, which can achieve the following technical effects: A. By standardizing and structuring the gridded data of numerical forecasts, an automated conversion link from numerical forecast products to operational text is established, eliminating the manual interpretation link and significantly improving the report production efficiency; B. By constructing a multi-source domain knowledge base and constraining the generation process of the large language model through context injection, the output content is ensured to conform to meteorological professional standards, improving the professionalism and accuracy of the report; C. By generating multiple candidate reports in parallel and driving the large language model to automatically score and select the best report, the randomness problem of single generation is fundamentally overcome, ensuring the quality and stability of operational output. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of an automated meteorological report generation system based on a large language model, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the overall system architecture according to an embodiment of the present invention; Figure 3 This is a flowchart of the numerical weather prediction data processing module according to an embodiment of the present invention; Figure 4 This is a hierarchical prompt word structure diagram according to an embodiment of the present invention; Figure 5 This is a flowchart of the multi-candidate parallel generation and automatic scoring and selection process according to an embodiment of the present invention; Figure 6 This is a flowchart of the automated meteorological report generation method based on a large language model according to an embodiment of the present invention. Detailed Implementation

[0018] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this document.

[0019] System Implementation Examples According to embodiments of the present invention, an automated meteorological report generation system based on a large language model is provided. Figure 1 This is a schematic diagram of an automated meteorological report generation system based on a large language model, according to an embodiment of the present invention. Figure 1 As shown, the automated meteorological report generation system based on a large language model according to an embodiment of the present invention specifically includes: The data preprocessing module 10 is used to acquire raw meteorological numerical forecast data and perform structured transformation on the raw meteorological numerical forecast data to generate standardized meteorological data in a preset format; specifically, it is used for: Acquire gridded forecast data output by numerical weather prediction models, perform quality checks on the gridded forecast data, and identify and correct outliers and missing data. Key meteorological elements in the repaired gridded forecast data are extracted and statistically analyzed to generate time-series statistical features. The time-series statistical features are converted into a structured text format to generate standardized meteorological data.

[0020] Knowledge base construction module 12 is used to build a multi-source meteorological professional knowledge base; specifically, it is used for: Acquire at least one type of meteorological professional information source to construct a multi-source meteorological professional domain knowledge base; wherein, the information source includes one or more of historical forecast reports, meteorological academic literature, and expert experience knowledge; the historical forecast reports include historical forecast texts, forecast terminology specifications, typical weather process description templates, and professional expression methods; the meteorological academic literature includes key conclusions, meteorological element threshold standards, and judgment rules in meteorological forecast-related research results; the expert experience knowledge includes industry judgment rules, regional climate characteristic descriptions, and typical forecast scenario descriptions; The information sources are decomposed into multiple independent knowledge entries. Each knowledge entry is labeled with a category tag, applicable scenario, and knowledge content, and stored in the multi-source meteorological professional domain knowledge base in a structured format. During system operation, the system selects either the first mode or the second mode based on preset conditions to provide knowledge content to the prompt word construction module. The first mode provides the full content of the multi-source meteorological professional knowledge base, while the second mode retrieves relevant subsets from the multi-source meteorological professional knowledge base based on current forecast characteristics and provides them.

[0021] The prompt word construction module 14, connected to the data preprocessing module and the knowledge base construction module, is used to receive the standardized meteorological data and the multi-source meteorological professional domain knowledge base, and configure them in prompt words at different levels, which are then concatenated to generate a unified model input command; specifically used for: Construct data-level prompts; wherein, the data-level prompts include the standardized meteorological data, task description, and output format requirements; Construct system-level prompt words; wherein, the system-level prompt words include the full content of the multi-source meteorological professional domain knowledge base, report format specifications, terminology usage specifications, and generation constraints; The system-level prompts and data-level prompts are concatenated in a preset order to obtain complete model input instructions.

[0022] Parallel generation module 16, connected to the prompt word construction module, is used to generate multiple candidate weather reports by parallel invoking a large language model based on the model input instructions and through differentiated generation parameters; specifically, it is used for: The model input command is used as the unified input. At least one differential generation parameter is set; wherein, the differential generation parameter includes one or more of the following: random sampling temperature parameter, Top-p parameter, Top-k parameter, or random seed parameter; Based on the differentiated generation parameters, the large language model is invoked in parallel to generate multiple candidate meteorological reports.

[0023] The preferred output module 18, connected to the parallel generation module, is used to automatically evaluate the quality of the multiple candidate meteorological reports and output the final meteorological report based on the evaluation results; specifically, it is used for: Construct scoring prompts; wherein, the scoring prompts include a description of the scoring task and multiple scoring dimensions; the scoring dimensions include forecast consistency, content completeness, language standardization, and format standardization; The large language model is invoked based on the scoring prompts to score the multiple candidate weather reports on each scoring dimension, and the comprehensive score of each candidate weather report is calculated based on the weight of each scoring dimension; wherein, the weight is a preset value or is adaptively adjusted according to the application scenario; The candidate meteorological report with the highest comprehensive score is selected as the final output meteorological report.

[0024] The system is also used for: Receive numerical weather forecast data for different target areas and output corresponding weather reports for those areas; The target area includes one or more of the following: wind farm area, air route area, or agricultural planting area.

[0025] The following describes in detail the above-mentioned technical solutions of the present invention with reference to the specific circumstances of the automated meteorological report generation system based on a large language model according to an embodiment of the present invention.

[0026] This invention provides an automated meteorological report generation system based on numerical model forecasting and large language model multi-candidate optimization. The system consists of five functional modules arranged collaboratively according to data flow, and the overall architecture is as follows: Figure 2 As shown, the five modules are: numerical weather prediction data processing module, multi-source domain knowledge base module, hierarchical prompt word construction module, multi-candidate parallel generation module, and automatic scoring, selection and distribution module. The system comprises five modules: Module 1 (Numerical Prediction Data Processing Module) receives real-time FGOALS-f2 S2S forecast grid data, performs quality control, feature extraction, and JSON structured conversion, and outputs structured forecast data; Module 2 (Multi-Source Domain Knowledge Base Module) manages a professional knowledge base consisting of historical forecast reports, meteorological academic literature, and expert experience knowledge, and outputs domain knowledge text; Module 3 (Hierarchical Prompt Term Construction Module) injects the structured data from Module 1 into data-level prompt terms and embeds the knowledge base content from Module 2 into system-level prompt terms, completing the two-level prompt term concatenation; Module 4 (Multi-Candidate Parallel Generation Module) uses the complete prompt terms output from Module 3 as input and calls the large language model API in parallel through differentiated Temperature parameters to generate N candidate reports; and Module 5 (Automatic Scoring, Optimization, and Distribution Module) constructs scoring prompt terms, drives the large language model to perform multi-dimensional weighted scoring on all candidate reports, selects the report with the highest comprehensive score, and automatically distributes it. The data flow is as follows: Module 1 (data processing) and Module 2 (knowledge base) run independently and in parallel, with their outputs converging in Module 3 (prompt word construction); Modules 3 to 5 form a unidirectional serial link, ultimately outputting a quality-optimized weather forecast report. Details are as follows: Module 1: Numerical Prediction Data Processing Module This module converts the raw grid data output by the FGOALS-f2 numerical weather prediction model into structured text data that can be directly read by large language models, such as... Figure 3 As shown in the diagram, the steps are executed sequentially, and the final output is JSON structured weather forecast data that can be directly used for prompt word injection. The specific implementation steps are as follows: The first step is data acquisition. The system automatically triggers at a fixed time each day to read the gridded forecast field of the target area (e.g., region A) from the output file of the FGOALS-f2 sub-seasonal-seasonal (S2S) real-time forecast product. The physical quantities read include key meteorological elements such as: 850hPa / 500hPa / 200hPa zonal winds (ua850 / ua500 / ua200), 850hPa / 500hPa / 200hPa meridional winds (va850 / va500 / va200), 850hPa / 500hPa / 200hPa geopotential height (zg850 / zg500 / zg200), sea level pressure (PSL), 2-meter air temperature (TREFHT), 10-meter wind speed (W10m), precipitation (PRECT), and the El Niño index. The forecast lead time covers daily forecast data for the next six months.

[0027] The second step is quality control. A feature-by-feature quality check is performed on the raw data: outliers exceeding physically reasonable ranges are identified and marked, such as TREFHT exceeding the regional historical extreme threshold, negative W10m values, and PSL exceeding the 900-1100 hPa range. Adjacent grid interpolation is used to correct these outliers. Missing data is filled using time-linear interpolation to ensure the integrity and reliability of the data used in subsequent processing.

[0028] The third step is the extraction and statistical analysis of key elements. For the target area, spatial averaging or target point interpolation is performed on the quality-controlled grid data to extract the following statistics: daily average, daily maximum, and daily minimum values ​​for each element; 10-meter wind speed is directly obtained from the W10m variable; wind speeds at 850hPa / 500hPa / 200hPa levels are synthesized from the corresponding U and V components, and the prevailing wind direction at each level is calculated; precipitation is calculated as a cumulative value over the time period. These statistics are then organized into a structured data table based on the time series.

[0029] The fourth step is structured format conversion. The statistical data obtained in the third step is converted into structured text in JSON format. An example of field design is shown below: {"date": "2025-01-01", "TREFHT_mean": 5.2, "TREFHT_max": 9.1, "TREFHT_min": 1.3, "W10m_mean": 4.5, "wind_direction_10m": "Northwest", "PSL_mean": 1015.3, "zg500_mean": 5820.0, "ua850_mean": -2.1, "va850_mean": 3.5, "PRECT_sum": 0.5, "nino_index": 0.3} This JSON structured data is output from this module and passed to the data-level prompts in module three (the hierarchical prompt construction module).

[0030] Module 2: Multi-Source Domain Knowledge Base Module This module's function is to build and maintain a professional knowledge base geared towards operational weather forecasting scenarios, providing domain knowledge references for the large language model during the report generation phase. The specific implementation is as follows: The knowledge base consists of content from three sources: the first is historical meteorological forecast reports, which collect and organize publicly released historical forecast texts from authoritative meteorological institutions, extracting forecast terminology standards, typical weather process description templates, and professional expressions; the second is meteorological academic literature, which extracts key conclusions, meteorological element threshold standards, and judgment rules from research results related to weather forecasting; and the third is expert experience knowledge, which includes industry judgment rules, descriptions of regional climate characteristics, and explanations of typical forecast scenarios manually entered by meteorological professionals.

[0031] The above three types of content are organized and stored in a knowledge base file in JSON format. Each knowledge entry contains three fields: category label, applicable scenario, and knowledge content.

[0032] During the report generation phase, the entire content of the knowledge base file is read and embedded as a string at a designated location in the system-level prompt words, serving as the reference context when the large language model generates the report. When the total number of tokens in the knowledge base does not exceed 50% of the model's context window limit (to reserve space for forecast data and generated content), this full injection method is used to ensure the model fully references all professional knowledge. When the knowledge base size exceeds this limit, it automatically switches to the retrieval-enhanced RAG generation method, retrieving the most relevant entries from the knowledge base based on the day's forecast characteristics and injecting the prompt words.

[0033] Module 3: Layered Prompt Keyword Construction Module This module's function is to assemble the structured forecast data from Module 1 and the domain knowledge base content from Module 2 into complete prompt words, which are then input into the large language model. The prompt words employ a two-level hierarchical structure, such as... Figure 4 As shown: System Prompt: The content is fixed and includes the following elements: format specifications for weather forecast reports (title, timeliness, and body structure requirements); professional terminology usage specifications (requiring the use of meteorological industry standard terms); all content of the Module 2 knowledge base (as a domain knowledge reference); generation constraints (such as report word count range, prohibition of conclusions that contradict numerical data, etc.).

[0034] User Prompt (Data-level Prompt): Content is dynamically generated daily and includes the following elements: the structured JSON forecast data for the day output by Module 1; a description of the report generation task (target timeframe, target area, key focus elements); and output format requirements (the model must strictly adhere to the format specified by the system-level prompt).

[0035] The two-level prompts are concatenated before each call to the large language model to form a complete input instruction, which is then passed to Module 4. That is, the system-level prompts in this module remain fixed, while the data-level prompts are updated daily, with the two levels concatenated before each call to form a unified input.

[0036] Module 4: Multi-candidate parallel generation module This module's function is to generate multiple candidate reports in parallel based on the prompt words constructed in Module 3. The specific implementation is as follows (see also Module 5). Figure 5 ): The system uses the complete prompt words output from Module 3 as unified input. It initiates N parallel call requests via a locally deployed large language model API (using the Dify framework with the Qwen3-32B model, deployed locally on an NVIDIA A100 server, and invoked through its REST API interface), where N is a positive integer greater than or equal to 10, flexibly set according to computing resources and timeliness requirements. Each call sets a different random sampling temperature parameter value in the API request parameters, ensuring that each generated result produces differentiated expressions while maintaining semantic consistency, thus covering multiple possible forms of model output.

[0037] After N calls are completed, all N candidate report texts are collected, stored in the candidate report list, and passed to Module 5.

[0038] Module 5: Automatic Scoring, Optimization, and Distribution Module This module automatically assesses the quality of all candidate reports, selects the best report, and distributes it. The detailed process is as follows: Figure 5 As shown, Figure 5 This demonstrates the complete backend process from prompt input to optimal report distribution, involving N parallel calls using differentiated Temperature parameters; after scoring, reports are sorted in descending order of weighted composite scores, and the optimal report is automatically selected, all without manual intervention. The specific implementation is as follows: The first step is to construct scoring prompts. The system constructs dedicated scoring prompts, including: a description of the scoring task (scoring candidate reports from multiple dimensions); definitions of scoring dimensions, totaling four dimensions: A. Forecast consistency, i.e., whether the forecast conclusions in the report are consistent with the input numerical forecast data, weight 40%; B. Content completeness, i.e. whether key meteorological elements (temperature, wind speed, wind direction, precipitation, air pressure, etc.) are all described, weight 30%; C. Language standardization, i.e. whether the language conforms to meteorological industry standards and whether the expression is fluent, weight 20%; D. Format standardization, i.e. whether the report structure conforms to the prescribed format, weight 10%; scoring requirements, requiring the model to give each report a score of 0-10 points according to the above dimensions, calculate a weighted comprehensive score, and output the scoring results in JSON format. The weights can also be adaptively adjusted according to actual conditions.

[0039] The second step is batch scoring. The candidate report list is combined with the scoring prompts, and a large language model is used to score all N candidate reports one by one, obtaining the comprehensive score for each report.

[0040] The third step is the selection of the best report. The system analyzes the scoring results, sorts them in descending order of overall score, and automatically selects the candidate report with the highest score as the final output report for the day.

[0041] The fourth step is automatic distribution. After formatting the optimal report according to the preset format (including title, date, and body), it is automatically pushed to the target users through the system's pre-configured distribution interface, without any manual intervention.

[0042] Taking the generation of a weather forecast report for a wind farm on January 1, 2025 as an example (this invention is also applicable to weather reporting scenarios such as aviation and agriculture), the complete execution process of this invention is explained as follows: On the same day, the system automatically starts. Module 1 reads gridded forecast data for the wind farm's location for the next six months from the FGOALS-f2 S2S real-time forecast product output file. After quality control and element extraction, it generates JSON structured data containing daily temperature, wind speed, wind direction, precipitation, and air pressure. Module 2 reads the domain knowledge base JSON file to obtain all professional knowledge content. Module 3 embeds the knowledge base content into system-level prompt words and injects the JSON forecast data into data-level prompt words to complete the prompt word concatenation. Module 4 uses the prompt words as input, sets the Temperature parameter to be evenly distributed within the range of 0.3 to 0.9, and calls the Qwen3-32B large language model API deployed locally in Dify in parallel to generate N candidate reports (N≥10). Module 5 constructs scoring prompt words, calls the large language model to score all candidate reports one by one, parses the scoring JSON results, selects the report with the highest comprehensive score, formats it, and automatically pushes it to the wind farm operators to complete the daily report production task. The entire process requires no manual intervention.

[0043] The key to the embodiments of the present invention lies in: 1. Numerical Prediction Data Structure Processing Method: A complete processing method for raw gridded data output by numerical prediction models, including quality control, extraction of key meteorological elements, and JSON format structure conversion, to achieve automated conversion from numerical prediction products to usable input for large language models.

[0044] 2. Construction of a multi-source domain knowledge base and a full-context injection mechanism: A domain knowledge base is constructed by integrating three sources: historical forecast reports, meteorological academic literature, and expert experience knowledge. The knowledge base content is injected as a whole into system-level prompt words each time it is generated.

[0045] 3. A hierarchical prompt word construction method for meteorological professional scenarios: a two-level hierarchical prompt word architecture design with system-level prompt words that solidify domain knowledge and generation specifications, and data-level prompt words that dynamically inject real-time forecast data.

[0046] 4. Multi-candidate parallel generation mechanism: Based on the same prompt word input, a method is used to generate N candidate reports in parallel by setting differentiated random sampling temperature parameters and calling a large language model.

[0047] 5. Multi-dimensional optimization mechanism based on large language model for automatic scoring: A method that uses forecast consistency, content completeness, language standardization, and format standardization as scoring dimensions to drive the large language model to perform weighted scoring on all candidate reports and automatically select the optimal output.

[0048] 6. A fully automated approach consisting of interconnected modules: an end-to-end complete approach from numerical forecast data input to automatic distribution of the optimal report, covering the entire chain of data processing, knowledge injection, prompt word construction, parallel generation, score optimization and automatic distribution.

[0049] Preferably, the FGOALS-f2 model is used as the numerical forecast data source in the example. However, the data processing method of the present invention is also applicable to the output data of other numerical forecast models, including but not limited to: ERA5 reanalysis data and real-time forecast products of the European Centre for Medium-Range Weather Forecasts (ECMWF), output data of the GFS global forecast system of the National Center for Environmental Prediction (NCEP) of the United States, output data of the GRAPES numerical forecast model of the China Meteorological Administration, and any other numerical forecast models that output standard grid data formats (such as NetCDF, GRIB2, etc.). Changes in the data source do not affect the implementation of the core method of the present invention.

[0050] Preferably, the example uses JSON format to store structured forecast data, but it can also be replaced with other structured text formats, including but not limited to: time-series data tables in CSV format, hierarchical data structures in XML format, and tabular text in Markdown format. All of the above formats can be used as input to a large language model to achieve the same data injection effect.

[0051] Preferably, the example uses the method of injecting the entire knowledge base content into the prompt words. When the knowledge base is large and exceeds the context window limit of the large language model, it can be replaced by the Retrieval-Augmented Generation (RAG) method: based on the characteristics of the daily forecast data (such as season, region, and main weather types), the most relevant knowledge entries are retrieved from the knowledge base, and only the retrieval results are injected into the prompt words. Both injection methods can achieve the goal of model generation constrained by domain knowledge, and the effect is equivalent.

[0052] Preferably, the example uses a localized deployment scheme based on the Dify framework, but it can also be replaced by the following methods: deployment using other local inference frameworks such as Ollam, vLLM, and FastChat; or, if data security conditions permit, calling commercial large language model services such as the GPT series and DeepSeek API through API calls. Changes in deployment methods do not affect the implementation of the multi-candidate generation and scoring optimization mechanism of this invention.

[0053] Preferably, in the example, N ≥ 10 candidate reports are set, and the specific value can be flexibly set according to computing resources and time requirements. The range of N is N ≥ 2 to achieve the basic function of multi-candidate selection. The larger the value of N, the higher the quality of the best candidate in theory, but the greater the corresponding consumption of computing resources.

[0054] Preferably, the example achieves differentiated generation of candidate reports by setting different random sampling temperature parameters. Alternatively, the following parameters can be used in combination or replaced: Top-p parameter (controls the cumulative probability range of words considered during sampling); Top-k parameter (controls the number of candidate words considered in each sampling step); and random seed parameter (generates differentiated output by setting different random seeds). The use of these parameters individually or in combination can achieve the goal of generating multiple differentiated candidate reports based on the same prompt word input.

[0055] Preferably, in the example, the large language model itself undertakes the scoring task, with four dimensions: forecast consistency, content completeness, language standardization, and format standardization. Alternative solutions may include: adding or adjusting scoring dimensions, such as adding dimensions like data citation accuracy and audience readability; employing a hybrid scoring mechanism combining rule-based scoring and model scoring, such as using keyword matching rules to check whether key meteorological elements are covered, and then combining this with model scoring for a comprehensive evaluation; or using another independent large language model (different from the generative model) to undertake the scoring task, in order to reduce self-assessment bias.

[0056] It is worth noting that the complete method of this invention can also be extended to other professional fields of report automatic generation scenarios besides meteorological forecast reports, such as: automatic generation of hydrological forecast reports based on hydrological numerical model output data, automatic generation of environmental quality forecast reports based on air quality numerical model output data, and automatic generation of aviation meteorological reports based on meteorological forecast data. The above scenarios are completely consistent with the method steps of this invention, differing only in the data source and the content of the domain knowledge base.

[0057] Method Implementation Examples According to embodiments of the present invention, a method for automatically generating meteorological reports based on a large language model is provided. Figure 6 This is a flowchart of the automated meteorological report generation method based on a large language model according to an embodiment of the present invention, such as... Figure 6 As shown, the automated meteorological report generation method based on a large language model according to an embodiment of the present invention specifically includes: Step S601: Obtain raw meteorological numerical forecast data through the data preprocessing module, and perform structured transformation on the raw meteorological numerical forecast data to generate standardized meteorological data in a preset format; Step S602: Construct a multi-source meteorological professional domain knowledge base through the knowledge base construction module; Step S603: The standardized meteorological data and the multi-source meteorological professional domain knowledge base are received through the prompt word construction module and configured into prompt words at different levels. After being spliced ​​together, a unified model input instruction is generated. Step S604: Based on the model input instructions, the parallel generation module calls the large language model in parallel through differentiated generation parameters to generate multiple candidate meteorological reports. Step S605: The multiple candidate meteorological reports are automatically evaluated by the preferred output module, and the final meteorological report is output based on the evaluation results.

[0058] The embodiments of the present invention are method embodiments corresponding to the system embodiments described above. The specific operations of each step can be understood by referring to the description of the system embodiments, and will not be repeated here.

[0059] In summary, compared with the prior art, the embodiments of the present invention have the following beneficial effects: 1. Solved the problem of automated conversion of numerical weather prediction data into operational text: Existing technologies lack a systematic processing flow for gridded data and rely on manual interpretation; the embodiments of this invention realize fully automated processing of numerical weather prediction data through a complete processing chain of quality control, feature extraction and JSON structured conversion, eliminating the manual interpretation link and greatly improving report production efficiency.

[0060] 2. Significantly improves the professionalism and accuracy of generated reports: Existing technologies directly call general-purpose large language models, which, due to the lack of domain knowledge constraints, are prone to generating forecast conclusions that are inconsistent with numerical data; the embodiments of this invention, through full context injection from a multi-source domain knowledge base and hierarchical prompt word constraints, ensure that the generated content conforms to meteorological industry standards and that the forecast conclusions are consistent with the numerical data.

[0061] 3. It fundamentally solves the problem of unstable quality caused by the randomness of the output of large language models: existing technologies generate and output directly in a single step, and the quality depends entirely on the random sampling results of the model; the embodiment of this invention generates N candidate reports in parallel and performs multi-dimensional automatic scoring and selection to ensure the quality of the final output in a probability coverage manner, thus meeting the stability requirements of business applications.

[0062] 4. Achieves fully automated production without human intervention: Existing technologies require human intervention in data processing, report writing, quality review, and other aspects; the embodiments of this invention automate all of the above aspects, forming a complete end-to-end automated chain from numerical forecast data input to optimal report distribution and output.

[0063] 5. High data security: The embodiments of this invention adopt a localized deployment scheme for large language models. The forecast data is processed on the local server throughout the process and is not transmitted to the external network, which meets the requirements for meteorological data security management.

[0064] Device Example 1 This invention provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, it performs the steps described in the method embodiment.

[0065] Device Example 2 This invention provides a computer-readable storage medium storing an information transmission implementation program, which, when executed by a processor, performs the steps described in the method embodiment.

[0066] The computer-readable storage media described in this embodiment include, but are not limited to, ROM, RAM, disk, or optical disk.

[0067] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for automatically generating meteorological reports based on a large language model, characterized in that, include: The data preprocessing module is used to acquire raw meteorological numerical forecast data and perform structured transformation on the raw meteorological numerical forecast data to generate standardized meteorological data in a preset format. The knowledge base construction module is used to build a multi-source meteorological professional knowledge base; The prompt word construction module, connected to the data preprocessing module and the knowledge base construction module, is used to receive the standardized meteorological data and the multi-source meteorological professional domain knowledge base, and configure them in prompt words at different levels, which are then concatenated to generate a unified model input instruction; The parallel generation module, connected to the prompt word construction module, is used to generate multiple candidate weather reports by calling the large language model in parallel through differentiated generation parameters based on the model input instructions. The preferred output module, connected to the parallel generation module, is used to perform automated quality assessment on the multiple candidate meteorological reports and output the final meteorological report based on the assessment results.

2. The system according to claim 1, characterized in that, The data preprocessing module is specifically used for: Acquire gridded forecast data output by numerical weather prediction models, perform quality checks on the gridded forecast data, and identify and correct outliers and missing data. Key meteorological elements in the repaired gridded forecast data are extracted and statistically analyzed to generate time-series statistical features. The time-series statistical features are converted into a structured text format to generate standardized meteorological data.

3. The system according to claim 1, characterized in that, The knowledge base construction module is specifically used for: Acquire at least one type of meteorological professional information source to construct a multi-source meteorological professional domain knowledge base; wherein, the information source includes one or more of historical forecast reports, meteorological academic literature, and expert experience knowledge; the historical forecast reports include historical forecast texts, forecast terminology specifications, typical weather process description templates, and professional expression methods; the meteorological academic literature includes key conclusions, meteorological element threshold standards, and judgment rules in meteorological forecast-related research results; the expert experience knowledge includes industry judgment rules, regional climate characteristic descriptions, and typical forecast scenario descriptions; The information sources are decomposed into multiple independent knowledge entries. Each knowledge entry is labeled with a category tag, applicable scenario, and knowledge content, and stored in the multi-source meteorological professional domain knowledge base in a structured format. During system operation, the system selects either the first mode or the second mode based on preset conditions to provide knowledge content to the prompt word construction module. The first mode provides the full content of the multi-source meteorological professional knowledge base, while the second mode retrieves relevant subsets from the multi-source meteorological professional knowledge base based on current forecast characteristics and provides them.

4. The system according to claim 1, characterized in that, The prompt word construction module is specifically used for: Construct data-level prompts; wherein, the data-level prompts include the standardized meteorological data, task description, and output format requirements; Construct system-level prompt words; wherein, the system-level prompt words include the full content of the multi-source meteorological professional domain knowledge base, report format specifications, terminology usage specifications, and generation constraints; The system-level prompts and data-level prompts are concatenated in a preset order to obtain complete model input instructions.

5. The system according to claim 1, characterized in that, The parallel generation module is specifically used for: The model input command is used as the unified input. At least one differential generation parameter is set; wherein, the differential generation parameter includes one or more of the following: random sampling temperature parameter, Top-p parameter, Top-k parameter, or random seed parameter; Based on the differentiated generation parameters, the large language model is invoked in parallel to generate multiple candidate meteorological reports.

6. The system according to claim 1, characterized in that, The preferred output module is specifically used for: Construct scoring prompts; wherein, the scoring prompts include a description of the scoring task and multiple scoring dimensions; the scoring dimensions include forecast consistency, content completeness, language standardization, and format standardization; The large language model is invoked based on the scoring prompts to score the multiple candidate weather reports on each scoring dimension, and the comprehensive score of each candidate weather report is calculated based on the weight of each scoring dimension; wherein, the weight is a preset value or is adaptively adjusted according to the application scenario; The candidate meteorological report with the highest comprehensive score is selected as the final output meteorological report.

7. The system according to claim 1, characterized in that, The system is also used for: Receive numerical forecast data for different target areas and output meteorological reports for the corresponding areas; The target area includes one or more of the following: wind farm area, air route area, or agricultural planting area.

8. A method for automatically generating meteorological reports based on a large language model, characterized in that, include: The raw meteorological numerical forecast data is obtained through the data preprocessing module, and the raw meteorological numerical forecast data is structurally transformed to generate standardized meteorological data in a preset format. A multi-source meteorological professional knowledge base is constructed using a knowledge base construction module; The standardized meteorological data and the multi-source meteorological professional domain knowledge base are received by the prompt word construction module and configured into prompt words at different levels. After being concatenated, a unified model input command is generated. Based on the input instructions of the model, the parallel generation module calls the large language model in parallel through differentiated generation parameters to generate multiple candidate weather reports. The multiple candidate meteorological reports are automatically evaluated using the optimized output module, and the final meteorological report is output based on the evaluation results.

9. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the method for automatically generating weather reports based on a large language model as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores an information transmission implementation program, which, when executed by a processor, implements the steps of the automated meteorological report generation method based on a large language model as described in claim 8.

Images