Large language model enhanced short-to-medium term hydrological prediction human-machine collaboration method and system

Abstract

This invention provides a human-machine collaborative method and system for short- and medium-term hydrological forecasting enhanced by a large language model, relating to the fields of hydrological forecasting and artificial intelligence. The method includes the following steps: constructing a hydrological forecasting knowledge base for the target watershed; obtaining natural language forecasting requirements and generating structured forecasting task descriptions; recommending solutions and completing parameters based on the knowledge base; executing business rule constraints and verifying solutions; calling the short- and medium-term hydrological forecasting system to perform forecast calculations; and intelligently interpreting forecast results and accumulating experience. The system includes six modules: user interaction, large language model interaction, hydrological forecasting knowledge base, business rule engine, system interface calls, data storage, and logs. Without changing the core architecture of existing short- and medium-term hydrological forecasting systems, this invention introduces a large language model to achieve semantic parsing of forecasting tasks, intelligent construction of forecasting solutions, and intelligent interpretation of forecast results, lowering the system's usage threshold and improving the interpretability of forecast results and the structured reusability of forecasting experience.

Description

Technical Field

[0001] This invention belongs to the field of hydrological forecasting and artificial intelligence technology, specifically involving a human-machine collaborative method and system for short- and medium-term hydrological forecasting enhanced by a large language model, applicable to business scenarios such as reservoir scheduling, hydrological information forecasting, and optimal allocation of water resources. Background Technology

[0002] Short- and medium-term hydrological forecasting is a crucial foundation for joint reservoir operation, flood control and drought relief, water resource allocation, and power production planning. In practice, short- and medium-term hydrological forecasting is typically conducted using a system developed based on web technologies such as Spring Boot and MyBatis. This system usually integrates functions such as meteorological and hydrological data querying, hydrological model calculation, forecast scheme configuration and management, and result comparison and analysis. It incorporates commonly used hydrological models such as the Xin'anjiang model, the Muskingen model, and hydrodynamic models, enabling forecasting of flow or water level processes for different forecast periods and supporting multi-scheme calculations and result comparison and analysis. However, as forecast objects become increasingly complex and operational precision increases, the limitations of such systems in supporting forecasting operations are gradually becoming apparent, mainly in the following aspects: (1) The system has a high barrier to entry. The short- and medium-term hydrological forecasting system involves multiple models, multiple parameters and multiple data sources. Users need to master the hydrological characteristics of the watershed, the principles of the model algorithm, the actual meaning of the parameters and the system operation process. New users find it difficult to master the configuration of multiple schemes and comprehensive analysis in a short period of time, which restricts the promotion and application of the system.

[0003] (2) Forecasting experience is difficult to structure and reuse. A large number of typical rain and flood processes, parameter combinations, scheme configurations and consultation conclusions accumulated in long-term forecasting practice exist in the form of reports, tables or personal experience. They are not deeply integrated with the calculation process of the forecasting system, making it difficult to reuse them among different personnel and different watersheds. It is difficult to form a unified scheme library and knowledge base.

[0004] (3) The existing system architecture is relatively rigid, making it difficult to directly introduce new intelligent interactive functions. Existing short- and medium-term hydrological forecasting systems have usually been running in production environments for a long time. Their internal model organization structure, interface specifications, and data flow are basically fixed. Significantly modifying the core modules poses stability risks and high implementation costs, which is not conducive to the rapid introduction of next-generation artificial intelligence technologies.

[0005] Large language models have demonstrated strong capabilities in natural language understanding, complex task decomposition, and knowledge retrieval and reasoning. Existing research primarily applies large language models to general question answering or report generation, lacking mature technical solutions for deep integration of large language models with short- and medium-term hydrological forecasting operational processes. In particular, a systematic solution has yet to be found for a human-machine collaborative mechanism to achieve semantic parsing of forecast tasks, intelligent construction of forecast schemes, and intelligent interpretation of forecast results without altering the existing forecasting system architecture. Meanwhile, existing research on applying machine learning or deep learning models to hydrological forecasting largely focuses on replacing or optimizing physical models, paying less attention to the human-machine collaborative issues in forecast task configuration and result interpretation, and has not yet formed a human-machine collaborative forecasting framework centered on large language models. Summary of the Invention

[0006] The purpose of this invention is to address the problems of high usage threshold, difficulty in reusing experience, and difficulty in quickly introducing intelligent interactive functions in existing short- and medium-term hydrological forecasting systems. It provides a human-computer collaborative method, system equipment, and storage medium for short- and medium-term hydrological forecasting enhanced by a large language model. Without changing the core architecture of the existing short- and medium-term hydrological forecasting system, it adds components such as large language model interaction and a knowledge base to achieve semantic parsing of forecasting tasks, intelligent construction of forecasting schemes, and intelligent interpretation of forecasting results. It also supports the structured accumulation and reuse of forecasting experience.

[0007] In order to achieve the above-mentioned objectives, the present invention: The first aspect provides a human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model, including the following steps: S1. Construct a hydrological forecast knowledge base for the target watershed, collect, structure, and vectorize historical forecast data for the target watershed, and establish a vector retrieval index. S2: Obtain the natural language forecast requirements input by forecasters, and after preprocessing, use a large language model to perform semantic parsing and feature extraction to generate a machine-processable structured forecast task description. S3, based on the structured forecast task description, retrieve similar historical forecast scenarios from the hydrological forecast knowledge base, generate candidate forecast schemes, and complete the automatic completion and adjustment of key parameters; S4. Input the candidate forecasting scheme into the business rule engine for rule verification. If it fails, the large language model combines the verification feedback information to iteratively correct it until a target forecasting scheme that meets the business rule constraints is generated. S5, convert the target forecast scheme into a call request that conforms to the existing short- and medium-term hydrological forecasting system interface specifications, call the short- and medium-term hydrological forecasting system to perform forecast calculations, and obtain forecast results and multi-scheme comparison evaluation indicators; S6. The large language model performs a comprehensive analysis of the forecast results and the multi-scheme comparative evaluation indicators to generate natural language interpretation results. At the same time, the structured information of the entire forecast process is written back to the hydrological forecast knowledge base to realize the accumulation and reuse of forecast experience.

[0008] Preferably, step S1 specifically includes: S11, determine the target watershed and its control reservoirs, control sections and hydrological stations, and collect at least one original operational data from the following related to the target watershed: historical hydrological forecasting schemes, typical rain and flood cases, scheduling procedures, model parameter configurations, forecast evaluation results and consultation records. S12, perform data cleaning and structured processing on the original business data, and parse the information on time, spatial location, watershed name, reservoir name, forecast type, flood type, model type and evaluation index into structured fields; S13, Perform feature extraction and vectorization processing on text data in structured materials, including at least word segmentation, entity recognition, keyword extraction and vector encoding, to obtain text vectors that can be used for similarity retrieval; S14. Based on the obtained text vector, establish a vector retrieval index, associate and store the structured fields with the corresponding text vectors, and form a hydrological forecast knowledge base oriented towards the target watershed.

[0009] Preferably, step S2 specifically includes: S21, receive natural language forecasting requests input by forecasters through a human-computer interaction interface. The natural language forecasting requests include at least one of the following: target reservoir or control section, forecast period, key indicators, and model preferences. S22, preprocess the natural language prediction requirements, including noise removal, unifying time and space representation, parsing technical terms, and obtaining standardized input text; S23, the standardized input text is input into the large language model, and the large language model performs semantic parsing and element extraction on the natural language forecast requirements to generate a structured forecast task description. The structured forecast task description is represented using a preset data structure and includes at least one target object identification field, forecast period configuration field, model combination field, input data source field, and output indicator field.

[0010] Preferably, step S3 specifically includes: S31, Based on the structured forecast task description, extract the target watershed or reservoir identifier, forecast type, flood type, forecast period range, and model type preference search conditions; S32, using the large language model to access the hydrological forecast knowledge base and its vector retrieval index, and based on the retrieval conditions and text vector similarity, retrieve at least one historical forecast scheme and parameter combination that is similar to the current forecast scenario in terms of watershed, reservoir, flood type and model type; S33, based on the similar historical schemes, at least one candidate forecast scheme is generated by combining the large language model with the structured forecast task description. The candidate forecast scheme includes at least one of the following: forecast model type, model combination method, initial values ​​of key parameters, areal rainfall weight, and forecast period configuration. S34, complete and adjust the missing or incomplete parameters in the candidate forecast schemes so that each candidate forecast scheme has a complete parameter configuration that can be directly executed in the short- and medium-term hydrological forecasting system.

[0011] Preferably, step S4 specifically includes: S41, The candidate forecast scheme is input into the business rule engine. The business rule engine obtains the corresponding rule configuration from the rule configuration library according to the identifier of the target reservoir or hydrological station. The rule configuration includes at least one or more of the following: forecast period boundary, model application scope, water balance constraint, watershed and reservoir safety constraint, and system resource constraint. S42, the rule execution engine performs a rule check on the candidate forecast scheme according to the rule configuration, and determines whether the forecast period setting exceeds the allowable range, whether the selected model type and combination method are in the allowable list, whether the key parameters fall within the allowable value range, and whether the water balance and flood control safety constraints are met. S43, if the candidate prediction scheme fails all rule checks, a verification feedback information containing the rule items that are not satisfied and their reasons is generated, and the verification feedback information is provided to the large language model. The large language model adjusts the candidate prediction scheme under the constraints of the verification feedback information to obtain a new candidate prediction scheme. S44. Repeat steps S41 to S43 until a target forecasting scheme that meets the preset business rules is generated.

[0012] Preferably, step S5 specifically includes: S51, the target forecast scheme is parsed and mapped according to the interface specification of the short- and medium-term hydrological forecasting system to generate an interface call request containing model identifier, parameter group, forecast period, forecast step size and input data source configuration; S52, through the meteorological and hydrological data query module in the short-to-medium-term hydrological forecasting system, the rainfall, flow, water level and auxiliary factor data for the corresponding time period are automatically obtained based on the interface call request; S53, call the hydrological forecast calculation module of the existing short- and medium-term hydrological forecast system, complete the forecast calculation according to the target forecast scheme configuration, obtain the forecast result corresponding to the target forecast scheme, and generate multi-scheme comparison evaluation index when there are multiple candidate schemes.

[0013] Preferably, step S6 specifically includes: S61, the forecast results and the multi-scheme comparison evaluation indicators are sent to the big language model, and the big language model combines typical rain and flood cases and historical evaluation information in the hydrological forecast knowledge base to generate natural language interpretation results containing flood process characteristics descriptions, risk warnings and scheme optimization suggestions; S62, the description of this structured forecasting task, the target forecasting scheme, the forecasting results, the final adopted scheme and its evaluation indicators are stored in the hydrological forecasting knowledge base in a structured form, and the newly added text data are vectorized and indexed and updated to realize the continuous accumulation and iterative utilization of forecasting experience.

[0014] Preferably, the structured forecasting task description is represented using a preset data structure, which includes at least one of the following fields: The target object identification field is used to identify the target reservoir, control section, or hydrological station; The forecast period configuration field is used to indicate the start and end times of the forecast and the forecast period tiers; The model combination field is used to indicate the lumped hydrological model, semi-distributed hydrological model, distributed hydrological model, or combination of different types of hydrological models used. The input data source field indicates at least one of the following data sources: real-time station rainfall, meteorological grid products, numerical weather forecast rainfall, and historical flow. Output indicator field, used to indicate at least one of the following output indicators: peak flow, peak time, flow or water level process line, and cumulative runoff.

[0015] Preferably, the data in the hydrological forecast knowledge base carries at least one tag, including watershed tag, reservoir tag, forecast type tag, flood type tag, model type tag, and forecast effect tag. When the large language model performs the similar historical forecast scenario retrieval in step S3, it filters and sorts the candidate historical records based on the tags.

[0016] Preferably, the business rule engine includes a rule configuration library and a rule execution engine. The rule configuration library stores the upper limit of the forecast period for different watersheds or reservoirs, the allowed model types and model combinations, the allowed value range of key parameters, and the flood control safety boundary requirements. When the rule execution engine verifies the candidate forecast scheme, it queries the corresponding rule configuration based on the identifier of the target reservoir or hydrological station and performs rule checks.

[0017] Preferably, the natural language interpretation results generated by the large language model in step S6 include: a comparative explanation of the differences in peak flow, peak time and cumulative runoff of different forecast schemes; a risk warning that may exceed the design flow, flood limit water level or control water level; an explanation of the similarity between the current forecast process and typical historical rain and flood processes; and at least one of the following: a draft structured text report for forecast consultation or dispatch orders.

[0018] The second aspect of the present invention provides: A human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model is provided for executing the aforementioned human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model. The system includes: The user interaction module is used to receive natural language forecast requests input by forecasters and to display forecast results and natural language interpretations. The large language model interaction module is used to communicate with the large language model, perform semantic parsing on the natural language forecasting requirements, generate a structured forecasting task description, and intelligently interpret the forecasting results after obtaining the forecasting results to generate natural language interpretation results. The hydrological forecasting knowledge base module is used to store historical forecasting schemes, typical rain and flood cases, scheduling procedures, model parameter configurations and forecasting evaluation results. It performs vectorization processing on text data, establishes a vector retrieval index, and provides knowledge retrieval services for the large language model interaction module. The business rules engine module is used to verify the candidate prediction schemes generated by the large language model interaction module according to preset business rules, and output the target prediction scheme or verification feedback information that meets the business rules. The system interface call module is used to convert the target forecast scheme into an interface call request that can be recognized by the short- and medium-term hydrological forecast system, call the meteorological and hydrological data query module and the hydrological forecast calculation module in the short- and medium-term hydrological forecast system, and obtain forecast results and multi-scheme comparison evaluation indicators. The data storage and log module is used to store the natural language forecasting requirements, the structured forecasting task description, the target forecasting scheme, the forecasting results and the final adopted scheme, and to write at least a portion of the information into the hydrological forecasting knowledge base module.

[0019] Preferably, the large language model interaction module, hydrological forecast knowledge base module, business rule engine module, system interface call module, and data storage and log module are deployed in software form on a server or cloud platform, and the module functions are implemented by the processor executing computer program instructions.

[0020] Preferably, the system is loosely coupled to the existing short- and medium-term hydrological forecasting system, without requiring modification to the core architecture and model calculation module of the existing short- and medium-term hydrological forecasting system.

[0021] Preferably, the business rule engine module includes a rule configuration library and a rule execution engine. The rule configuration library stores differentiated business rules for different river basins and reservoirs, and the rule execution engine is used to match rules and perform scheme verification.

[0022] The third aspect of the present invention provides: A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform a human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model.

[0023] The fourth aspect of the present invention provides: A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model.

[0024] The present invention has the following beneficial effects: 1. This invention uses a large language model enhancement system to couple with the existing short- and medium-term hydrological forecasting system via an external connection. It does not require major modifications to the core architecture and model calculation modules of the original system, and can realize semantic parsing of forecasting tasks, intelligent construction of forecasting schemes, and intelligent interpretation of forecasting results. It has high engineering feasibility and compatibility.

[0025] 2. This invention analyzes natural language forecasting requirements through a large language model, automatically extracts elements such as target reservoirs or hydrological stations, forecast period range, model combinations, and input data sources, and combines them with a knowledge base to achieve scheme recommendation and parameter completion. This significantly reduces manual configuration operations in complex interfaces, lowers the threshold for system use, and helps new staff to start forecasting work as soon as possible.

[0026] 3. This invention uses a hydrological forecasting knowledge base module to store historical forecasting schemes, typical rain and flood cases, and consultation conclusions in a structured and vectorized manner, and continuously updates them during each human-machine collaboration process. This effectively realizes the structured accumulation and cross-scenario reuse of forecasting experience, and provides technical support for building a scalable forecasting scheme library and knowledge base.

[0027] 4. In multi-scheme application scenarios, this invention uses a large language model to comprehensively analyze and interpret the results of different forecast schemes, generating textual descriptions and risk warnings for consultation and scheduling, making forecast conclusions more intuitive and easier to understand, and enabling forecasters to quickly grasp key processes and potential risk points.

[0028] 5. In terms of engineering deployment, the method and system structure are relatively open, which facilitates the connection with various existing short- and medium-term hydrological forecasting systems and different types of hydrological models. The knowledge base content and business rules can be gradually expanded according to the characteristics of different river basins and reservoirs. Therefore, it can be applied to multiple river basins, multiple reservoirs and various scheduling business scenarios. Attached Figure Description

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0030] Figure 1 This is a flowchart of a human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model, as described in a specific embodiment of the present invention.

[0031] Figure 2 This is a flowchart illustrating the process of obtaining natural language forecasting requirements and generating a structured forecasting task description in a specific embodiment of the present invention.

[0032] Figure 3 This is a flowchart illustrating the intelligent interpretation and experience accumulation of forecast results in a specific embodiment of the present invention.

[0033] Figure 4 This is a schematic diagram of the overall architecture of the human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model, as described in a specific embodiment of the present invention.

[0034] Figure 5 This is a schematic diagram of the three-layer deployment structure of the human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model, as described in a specific embodiment of the present invention.

[0035] Figure 6 This is a schematic diagram of the short-to-medium-term hydrological forecasting system for the upper reaches of the Yangtze River in a specific embodiment of the present invention.

[0036] Figure 7 This is a schematic diagram illustrating the application of the human-machine collaborative method and system for short- and medium-term hydrological forecasting based on large language model enhancement in a specific embodiment of the present invention. Detailed Implementation

[0037] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0038] Example 1: A human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model is characterized by the following steps: Step S1: Construct a hydrological forecast knowledge base for the target watershed, which involves systematically collecting, organizing, and vectorizing historical forecast data for the target watershed to form a hydrological forecast knowledge base that can be retrieved on demand by a large language model; S11: Determine the target watershed and its control reservoirs, control sections and hydrological stations, and collect original operational data related to the target watershed, such as historical hydrological forecasting schemes, typical rain and flood cases, scheduling procedures, model parameter configurations, forecast evaluation results and consultation records; S12: Perform data cleaning and structured processing on the original business data, and parse information such as time, spatial location, watershed name, reservoir name, forecast type, flood type, model type and evaluation index into structured fields; S13: Perform feature extraction and vectorization processing on the text data in the structured data, including at least word segmentation, entity recognition, keyword extraction and vector encoding, to obtain text vectors that can be used for similarity retrieval; S14: Based on the obtained text vector, establish a vector retrieval index, associate and store the structured fields with the corresponding text vectors, and form a hydrological forecast knowledge base oriented towards the target watershed; Step S2: Obtain natural language forecasting requirements and generate structured forecasting task descriptions. This involves collecting natural language requirements from forecasters through human-computer interaction, and then using a large language model to perform semantic parsing and element extraction, transforming them into machine-processable structured task descriptions. S21: Receive natural language forecasting requests input by forecasters through a human-computer interaction interface. The natural language forecasting requests include at least one of the following: target reservoir or control section, forecast period, key indicators, and model preferences. S22: Preprocess the natural language prediction requirements, including noise removal, unifying time and space representation, parsing technical terms, etc., to obtain standardized input text; S23: Input the normalized input text into the large language model, and the large language model performs semantic parsing and element extraction on the natural language forecast requirements to generate a structured forecast task description. The structured forecast task description is represented by a preset data structure and includes at least one target object identification field, forecast period configuration field, model combination field, input data source field and output indicator field. Step S3: Recommend solutions and complete parameters based on the knowledge base. This involves using a large language model in conjunction with the knowledge base to retrieve similar historical scenarios, generating candidate forecast solutions that meet the current needs, and automatically completing and adjusting key parameters. S31: Based on the structured forecast task description, extract retrieval conditions such as target watershed or reservoir identifier, forecast type, flood type, forecast period range, and model type preference; S32: Using the large language model, access the hydrological forecast knowledge base and its vector retrieval index, and based on the retrieval conditions and text vector similarity, retrieve at least one historical forecast scheme and parameter combination that is similar to the current forecast scenario in terms of watershed, reservoir, flood type and model type; S33: Based on the similar historical schemes, at least one candidate forecast scheme is generated by combining the large language model with the structured forecast task description. The candidate forecast scheme includes at least one of the following: forecast model type, model combination method, initial values ​​of key parameters, areal rainfall weight, and forecast period configuration. S34: Complete and adjust the missing or incomplete parameters in the candidate forecast schemes so that each candidate forecast scheme has complete parameter configurations that can be directly executed in the short- and medium-term hydrological forecasting system. Step S4: Execute business rule constraints and scheme verification, that is, input the candidate forecast schemes into the business rule engine, verify them item by item according to preset rules such as forecast period, safety boundary and model applicability, and combine feedback from the large language model for iterative correction until the target forecast scheme that meets the business constraints is obtained. S41: Input the candidate forecast scheme into the business rule engine. The business rule engine obtains the corresponding rule configuration from the rule configuration library according to the identifier of the target reservoir or hydrological station. The rule configuration includes at least one or more of the following: forecast period boundary, model application scope, water balance constraint, watershed and reservoir safety constraint, and system resource constraint. S42: The rule execution engine performs a rule check on each candidate forecast scheme according to the rule configuration, and determines whether the forecast period setting exceeds the allowable range, whether the selected model type and combination method are in the allowable list, whether the key parameters fall within the allowable value range, and whether the water balance and flood control safety constraints are met. S43: If the candidate prediction scheme fails all rule checks, a verification feedback information containing the rule items that are not satisfied and their reasons is generated, and the verification feedback information is provided to the large language model. The large language model adjusts the candidate prediction scheme under the constraints of the verification feedback information to obtain a new candidate prediction scheme. S44: Repeat steps S41 to S43 until a target forecasting scheme that meets the preset business rules is generated. Step S5: Call the short- and medium-term hydrological forecasting system to perform forecast calculations, that is, convert the target forecast scheme into a call request that conforms to the existing short- and medium-term hydrological forecasting system interface specifications, automatically obtain meteorological and hydrological input data and trigger model calculations to obtain forecast results and multiple scheme evaluation indicators; S51: Parse and map the target forecast scheme according to the interface specification of the short- and medium-term hydrological forecasting system, and generate an interface call request containing model identifier, parameter group, forecast period, forecast step size and input data source configuration; S52: Through the meteorological and hydrological data query module in the short-to-medium-term hydrological forecasting system, the rainfall, flow rate, water level and other auxiliary factor data for the corresponding time period are automatically obtained based on the interface call request; S53: Call the hydrological forecast calculation module in the short-to-medium-term hydrological forecast system, complete the forecast calculation according to the target forecast scheme configuration, obtain the forecast result corresponding to the target forecast scheme, and generate a multi-scheme comparison evaluation index when there are multiple candidate schemes; Step S6: Perform intelligent interpretation and experience accumulation of forecast results. This involves using a large language model to comprehensively analyze the forecast results and multi-scheme comparison information, generate textual interpretations for consultation and scheduling, and write the information of the entire forecast process back to the knowledge base for continuous learning. S61: The forecast results and the multi-scheme comparison evaluation indicators are sent to the big language model. The big language model combines typical rain and flood cases and historical evaluation information in the hydrological forecast knowledge base to generate natural language interpretation results that include descriptions of flood process characteristics, risk warnings and suggestions for optimal scheme selection. S62: The structured forecast task description, the target forecast scheme, the forecast results, the final adopted scheme and its evaluation indicators are stored in the hydrological forecast knowledge base in a structured form, and the newly added text data are vectorized and indexed and updated to realize the continuous accumulation and iterative utilization of forecast experience.

[0039] Furthermore, the structured forecasting task description is represented using a preset data structure, which includes at least one of the following fields: (1) Target object identification field, used to identify the target reservoir, control section or hydrological station; (2) Forecast period configuration field, used to indicate the start and end times of the forecast and the forecast period classification; (3) Model combination field, used to indicate the lumped hydrological model, semi-distributed hydrological model, distributed hydrological model or combination of different types of hydrological models used; (4) Input data source field, which indicates at least one of the following data sources: real-time station rainfall, meteorological grid products, numerical weather forecast rainfall, and historical flow; (5) Output index field, used to indicate at least one of the following output indexes: peak flow, peak time, flow or water level process line and cumulative runoff.

[0040] Furthermore, the data in the hydrological forecast knowledge base carries at least one tag, including watershed tag, reservoir tag, forecast type tag, flood type tag, model type tag, and forecast effect tag. When the large language model performs the similar historical scheme retrieval in step S3, it filters and sorts the candidate historical records based on the tags.

[0041] Furthermore, the business rule engine includes a rule configuration library and a rule execution engine. The rule configuration library stores the upper limit of the forecast period for different watersheds or reservoirs, the allowed model types and model combinations, the allowed value range of key parameters, and the flood control safety boundary requirements. When the rule execution engine verifies the candidate forecast scheme, it queries the corresponding rule configuration based on the identifier of the target reservoir or hydrological station and performs rule checks.

[0042] Furthermore, the natural language interpretation results generated by the large language model in step S6 include: a comparative explanation of the differences in peak flow, peak time and cumulative runoff of different forecast schemes; a risk warning that may exceed the design flow, flood limit water level or control water level; an explanation of the similarity between the current forecast process and typical historical rain and flood processes; and at least one of the following: a draft structured text report for forecast consultation or dispatch orders.

[0043] Example 2: like Figure 1 As shown, a human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model can be implemented according to the following steps: Step S1: Construct a hydrological forecasting knowledge base.

[0044] The historical forecast schemes, typical rain and flood cases, dispatching procedures, model parameter configurations and forecast evaluation results are standardized, organized and stored in fields. Text data such as reports and consultation records are processed by word segmentation, noise reduction and vectorization. A vector retrieval index for calling large language models is established to form a hydrological forecast knowledge base.

[0045] Step S2: Obtain natural language forecasting requirements and generate a structured forecasting task description.

[0046] like Figure 2As shown, forecasters input their natural language forecasting requirements through the user interaction module, such as "Based on the latest rainfall forecast, please provide a 7-day inflow forecast for a certain reservoir and compare it with last week's forecast." The large language model interaction module sends this input and preset prompts to the large language model, which outputs a structured forecasting task description, extracting elements such as the target reservoir or hydrological station, the forecast period, the forecast model type or combination, the input data source type, and output indicators. In this embodiment of the invention, the structured forecasting task description can be represented using a preset data structure, which includes at least one of the following fields: (1) Target object identification field, used to identify the target reservoir, control section or hydrological station; (2) Forecast period configuration field, used to indicate the start and end times of the forecast and the forecast period classification; (3) Model combination field, used to indicate the lumped hydrological model, semi-distributed hydrological model, distributed hydrological model or combination of different types of hydrological models (e.g., Xin'anjiang model, Muskingen model, etc.). (4) Input data source field, which indicates at least one of the following data sources: real-time station rainfall, meteorological grid products, numerical weather forecast rainfall, and historical flow; (5) Output index field, used to indicate at least one of the following output indexes: peak flow, peak time, flow or water level process line and cumulative runoff; (6) Business type field, which indicates the business purpose of this forecast scenario, such as flood control scheduling, reservoir water storage or power production plan.

[0047] The above fields can be stored in key-value pairs or represented in JSON structure, object classes, or other formats. The large language model interaction module generates this data structure based on the natural language input, and the system interface call module maps the data structure fields to the interface parameters of the short- and medium-term hydrological forecasting system.

[0048] Step S3: Recommend solutions and complete parameters based on the knowledge base.

[0049] Based on the structured forecast task description, the large language model interaction module retrieves historical forecast tasks and scheme records similar to the current scenario from the hydrological forecast knowledge base module, obtains information such as the corresponding model combination, initial parameter values, areal rainfall weights, and forecast period configuration, recommends schemes and completes parameters for the current task, and generates one or more candidate forecast schemes.

[0050] Step S4: Execute business rule constraints and verify the solution.

[0051] The business rules engine module, based on the target reservoir or hydrological station identifier, reads the corresponding forecast period upper limit, allowed model types and combinations, key parameter value ranges, and flood control safety boundaries from the rule configuration library to perform rule checks on candidate forecast schemes. If some candidate schemes violate the rules, verification feedback information is generated and returned to the large language model interaction module. The large language model then adjusts the candidate scheme configuration under the constraints of the feedback information until a target forecast scheme that meets the business rules constraints is generated.

[0052] Step S5: Call the short- and medium-term hydrological forecasting system to perform forecast calculations.

[0053] The system interface call module converts the target forecast scheme into an interface call request that can be recognized by the short- and medium-term hydrological forecasting system. It automatically configures interface parameters such as watershed, forecast object, start and end time, model combination, and input data source. It calls the meteorological and hydrological data query submodule to obtain input data such as rainfall and flow, and calls the hydrological forecast calculation submodule to complete the forecast calculation. When multiple scheme calculations are supported, the scheme management and comparison submodule can be called to obtain the forecast results and evaluation indicators of different schemes.

[0054] Step S6: Intelligent interpretation of forecast results and accumulation of experience.

[0055] The system interface call module returns the forecast results and multi-scheme comparison evaluation indicators to the large language model interaction module. The large language model interaction module, combining historical cases and scheduling procedures from the hydrological forecasting knowledge base, generates a natural language interpretation result containing descriptions of flood process characteristics, risk warnings, and optimal scheme recommendations. It then structurally stores the description of this forecasting task, the target forecasting scheme, the forecast results, and the final adopted scheme into the hydrological forecasting knowledge base, enabling the continuous accumulation and reuse of experience. The specific implementation process is as follows: Figure 3 As shown.

[0056] like Figure 4 As shown, the human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model provided in this embodiment mainly includes: The user interaction module provides a natural language input and result display interface for forecasters. It sends the natural language forecasting requests input by forecasters to the large language model interaction module and displays the structured forecasting task description, forecasting results and natural language interpretation content in the form of text or charts. The large language model interaction module is used to communicate with the preset large language model service, perform semantic parsing on the natural language forecasting requests sent by the user interaction module, generate a structured forecasting task description, and analyze the forecasting results after obtaining them from the system interface call module and generate natural language interpretation content. The hydrological forecast knowledge base module is used to store information such as historical forecast tasks, forecast schemes, model parameters, forecast results, typical rain and flood cases, scheduling procedures and technical specifications. It embeds and vectorizes text data and stores it in a vector database, providing knowledge query services based on keyword retrieval and vector similarity retrieval for the large language model interaction module. The business rules engine module includes a rule configuration library and a rule execution engine. The rule configuration library stores rule entries for different watersheds, reservoirs, and business types, including allowed model types and combinations, upper limits of forecast periods, allowed value ranges for key parameters, and flood control safety boundary requirements. The rule execution engine performs rule checks on candidate forecast schemes based on the target object and business type, and outputs verification results and feedback information. The system interface call module is used to map the watershed, forecast object, start and end time, model combination and data source configuration information in the structured forecast task description and target forecast scheme to the interface parameters of the short- and medium-term hydrological forecast system. It calls the meteorological and hydrological data query submodule in the short- and medium-term hydrological forecast system to obtain historical flow and rainfall data, and then calls the hydrological forecast calculation submodule to perform forecast calculation. It can also call the scheme management and comparison submodule to obtain forecast results and evaluation indicators of multiple schemes. The data storage and log module is used to record the entire process of human-machine collaboration, including natural language forecasting requirements, structured forecasting task descriptions, candidate forecasting schemes and target forecasting schemes, operational rule verification results, forecasting calculation results, multi-scheme comparison and evaluation indicators, and the final adopted scheme. Representative records are synchronized to the hydrological forecasting knowledge base module to continuously enrich the knowledge base content.

[0057] The short-to-medium-term hydrological forecasting system is an existing operational system, typically including a meteorological and hydrological data query submodule, a hydrological forecast calculation submodule, and a scheme management and comparison submodule. This embodiment establishes a loosely coupled data and task interaction relationship with the short-to-medium-term hydrological forecasting system through a system interface call module, without altering the original system's core architecture and model calculation process.

[0058] like Figure 5 As shown, the human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model in this embodiment of the invention can adopt a three-layer deployment structure in engineering practice, including: (1) User layer: mainly includes forecasting personnel terminals, which can be PC workstations or mobile terminals. The user layer accesses the user interaction module through a browser or dedicated client to realize natural language input and result display.

[0059] (2) Application Layer: Deploy the large language model interaction module, hydrological forecast knowledge base module, business rule engine module, system interface call module, and data storage and log module, and integrate them with the existing short- and medium-term hydrological forecasting system. The application layer can be deployed on a data center server or cloud platform environment and communicate with the user layer via HTTP / HTTPS and other methods.

[0060] (3) Data Layer: This includes a basic hydrological and meteorological database, a model configuration library, a forecast results library, and a hydrological forecast knowledge base, used to store basic data, model configurations, running results, and knowledge base vector data. The application layer module interacts with the data layer through a unified data access interface to read and write data.

[0061] The three-layer deployment structure described above allows for the smooth introduction of a human-machine collaborative system enhanced by a large language model without altering the existing short- and medium-term hydrological forecasting system deployment method, and supports subsequent module expansion and performance upgrades.

[0062] Example 3: This embodiment uses the short-to-medium-term inflow forecast of the cascade reservoir group in the upper reaches of the Yangtze River as an example to illustrate the application process of the method of the present invention in a practical engineering system. For example... Figure 6 As shown, a short-to-medium-term hydrological forecasting system for the upper reaches of the Yangtze River based on the SpringBoot architecture has been built at the project site. This system integrates functions such as forecast calculation, scheme access, clearing interaction, result storage, rainfall display, forecast parameter viewing, and multi-scheme management. Without changing the core architecture of the system, this invention deploys a human-computer collaborative system enhanced by a large language model to realize intelligent forecast task configuration and result interpretation.

[0063] Step S1: Construct a hydrological forecasting knowledge base for the upper reaches of the Yangtze River. In this embodiment, the hydrological forecasting knowledge base module focuses on the upper reaches of the Yangtze River basin, selecting important processes of the main streams and tributaries such as the Jinsha River, Min River, and Jialing River over the past 10-15 years, including regional rainstorms and floods, autumn floods, and low-water replenishment processes. It collects corresponding forecasting task records, the parameter sets of the Xin'anjiang model, the parameters of the Muskingan model, model combination methods, forecast period settings, areal rainfall weighting schemes, and forecasting evaluation indicators, organizing them into structured records and storing them in the knowledge base. It also cleans and annotates documents such as annual consultation minutes of the upper reaches of the Yangtze River basin, dispatching plan descriptions, and forecasting and recapture reports, adding basin tags, control station tags, flood type tags, and operational type tags, and storing them in a vector database through embedding vectorization. Simultaneously, it summarizes engineering experience such as the runoff generation and confluence characteristics of the main streams and tributaries of the upper reaches of the Yangtze River, the correspondence between control stations and control sections, the applicable scope of typical models, and parameter sensitivity, forming knowledge entries on model applicability and the basis for operational rule configuration.

[0064] Step S2: Obtain engineering forecast requirements and generate a structured forecast task description. During the flood season of a certain year, when analyzing a new round of rainfall, the Yangtze River Upper Reaches Forecasting Center input an engineering forecast requirement through the user interaction module: "Please combine the latest numerical rainfall forecast to provide the inflow forecast for a certain control station on the main stream of the upper Yangtze River and two major tributary reservoirs for the next 10 days, focusing on whether there will be near-guaranteed flow or exceeding the flood limit water level in the next 5 days, and compare it with the previous forecast plan." The large language model interaction module sends this natural language request to the large language model, which combines the large language model with the knowledge base. The content undergoes semantic parsing, outputting a structured forecast task description. The target objects include the main stream control station A, tributary reservoir B, and tributary reservoir C in the upper reaches of the Yangtze River. The forecast period is 10 days, with a step size of 1 hour. The main stream uses the Xin'anjiang model combined with the Muskingen model, while the tributaries use the Xin'anjiang model combined with reservoir capacity-water level calculation. The input data sources are measured rainfall and gridded rainfall from numerical weather prediction. The output indicators are the inflow process line, peak flow, peak time, and cumulative inflow over the next 10 days. The business type is the coordinated scheduling of flood control and water storage in the upper reaches of the Yangtze River.

[0065] Step S3: Recommending solutions and completing parameters based on the upper reaches of the Yangtze River knowledge base. The large language model interaction module accesses the hydrological forecasting knowledge base and its vector retrieval index, searching for similar historical events according to conditions such as "upper reaches of the Yangtze River main stream + corresponding tributary + flood control during the flood season + 10-day forecast period". It obtains model combinations, parameter sets, areal rainfall weights, forecast period configurations, and corresponding forecast performance evaluation results for several typical events, and generates multiple candidate forecasting solutions accordingly. Examples include solutions referencing parameter combinations from a typical flood season rainstorm event, solutions using parameter sets that perform well under similar basin average rainfall conditions, and solutions that make minor adjustments to sensitive parameters based on solutions currently in use in the production system. The candidate solutions already include information such as model combinations, initial parameter values, areal rainfall weights, and forecast period configurations.

[0066] Step S4: Execute business rule constraints and scheme verification for the upper reaches of the Yangtze River. The business rule engine module reads business rule entries related to the main stream control station A and tributary reservoirs B and C of the upper Yangtze River from the rule configuration library. These include flood control water level and scheduling curve constraints, downstream control section guaranteed flow constraints, upper limits of the forecast period and output time step requirements for different time periods, parameter value range suggestions, and system parallel scheme number limits, etc., and verifies each candidate scheme one by one. If it is found that the forecast period configuration of a certain scheme exceeds the upper limit allowed for the current stage, or some parameter values ​​exceed the preset range, the specific violated rule entry is returned as verification feedback information to the large language model interaction module. The large language model automatically adjusts the forecast period, output step size, or some parameters of the scheme until all candidate schemes pass the rule verification, and finally determines a set of target forecast schemes.

[0067] Step S5: Call the Yangtze River Upper Reaches Short- and Medium-Term Hydrological Forecasting System to perform forecast calculations. The system interface calling module maps structured information such as basin scope, control stations, tributary reservoirs, forecast start time, forecast period configuration, model combination scheme, and input data source to the standard interface parameters of the existing Yangtze River Upper Reaches Short- and Medium-Term Hydrological Forecasting System based on the target forecast scheme set, and triggers the forecast calculation task through the preset interface. Upon receiving a request, the short-to-medium-term hydrological forecasting system automatically extracts measured rainfall data from rain gauge stations in the upper reaches of the Yangtze River, measured flow data from control station A, and real-time water levels and inflow data from tributary reservoirs B and C via the meteorological and hydrological data query module. It reads the latest numerical weather forecast gridded rainfall products and calculates the areal average rainfall according to the forecast zones. It then calls the Xin'anjiang model to calculate runoff generation for each zone and the Muskingen model to perform confluence calculations of the main stream and tributaries, as well as main stream channel calculations. This generates hourly inflow forecasts for control station A and tributary reservoirs B and C for the next 10 days. When multiple target forecasting schemes are running simultaneously, the scheme management and comparison module outputs the forecast results of each scheme and multi-scheme comparison evaluation indicators, including peak deviation, peak time deviation, process correlation coefficient, and Nash efficiency, and returns them to the system interface calling module.

[0068] Step S6: Intelligent Interpretation of Forecast Results and Accumulation of Engineering Experience. The system interface module organizes the acquired forecast results from various schemes and the comparative evaluation indicators of multiple schemes into a unified structured format and sends it to the large language model interaction module. When calling the large language model service, the large language model interaction module includes relevant knowledge base entries from the upper reaches of the Yangtze River. Combining the current calculation results with historical typical process information, the large language model automatically generates natural language interpretations for engineering consultations. These interpretations include a description of the overall trend and key periods of inflow into control station A and tributary reservoirs B and C over the next 10 days; warnings about risks such as approaching or exceeding guaranteed flow and flood control limits within the next 5 days; a difference analysis of different schemes in terms of peak size, peak time, total runoff, and overall fit; reasons for recommending the preferred scheme; and a qualitative explanation of the similarity between this process and historical typical flood processes in the upper reaches of the Yangtze River. The above natural language interpretation results, along with charts, are displayed to the on-duty forecasters and consultation personnel through the user interaction module and can be directly used as a consultation presentation outline or a draft of the forecast analysis report. After the consultation, this embodiment will write back the structured information such as the description of the structured forecasting task, the set of target forecasting schemes, the forecasting results of multiple schemes, the comparative evaluation indicators of multiple schemes, and the final adopted scheme and scheduling conclusions into the hydrological forecasting knowledge base, so as to realize the rolling accumulation and reuse of engineering experience in the upper reaches of the Yangtze River.

[0069] Example 4: This embodiment takes the short-to-medium-term inflow forecast of a single flood control reservoir as the application object. The existing short-to-medium-term hydrological forecasting system is a localized system developed based on the Spring Boot architecture, which has built-in Xin'anjiang model and reservoir calculation model. The method and system of this invention are loosely coupled to the existing system. The specific implementation steps are as follows: S1. Construct a hydrological forecasting knowledge base for the reservoir: The control basin, hydrological stations, and monitoring sections of the reservoir were determined. Historical forecast schemes, rainstorm and flood cases, reservoir operation procedures, Xin'anjiang model parameter configurations, forecast evaluation results, and flood control consultation records from the past 10 years were collected. The above data were cleaned to extract structured fields such as forecast time, flood type, model parameters, and peak flow. Text data such as consultation records and operation procedures were segmented, entity recognized, and vector encoded to generate text vectors. A vector retrieval index was built based on the text vectors, and reservoir tags, flood type tags, and forecast effect tags were added to the knowledge base data. The structured fields were associated and stored with the text vectors to form a dedicated hydrological forecast knowledge base.

[0070] S2. Obtain natural language forecasting requirements and generate a structured forecasting task description: Forecasters input their natural language requests through the user interaction module: "Please combine the latest numerical rainfall forecast to provide a 7-day inflow forecast for this reservoir, paying particular attention to whether it exceeds the flood limit level, and output the peak flow and peak time." The system preprocesses the requirement, removing noise and parsing technical terms, and inputs the standardized text into a large language model. The large language model performs semantic parsing and feature extraction to generate a structured forecast task description, represented in JSON format, specifically: { "Target object": "Inlet section of XX Reservoir" "Forecast period": "February 8, 2026 to February 14, 2026, 1-hour step". "Model combination": "Xin'anjiang model + reservoir calculation model" "Input data source": "Numerical weather forecast rainfall + measured reservoir flow" "Output Indicators": "Inflow rate curve, peak flow, peak occurrence time, and flood control level warning." } S3. Recommended Scheme and Parameter Completion: Based on the structured task description, the large language model retrieves similar historical scenarios of rainstorm and flood types and 7-day forecast periods from the knowledge base, and obtains the model parameters and areal rainfall weight configurations of 3 sets of optimal historical forecast schemes. Based on similar schemes, it generates 3 sets of candidate forecast schemes, automatically fills in the missing Xin'anjiang model runoff generation parameters, confluence parameters and areal rainfall weight coefficients in the candidate schemes, so that all candidate schemes have complete executable parameters.

[0071] S4. Business rule verification and scheme modification: The business rules engine retrieves the reservoir's specific rules from the rule configuration library: a maximum forecast period of 10 days, a range of parameters for the Xin'anjiang model, and a flood control constraint of a flood limit water level of 105m. The rule execution engine validated the three candidate schemes and found that the penetration coefficient of the Xin'anjiang model in one scheme exceeded the range. The engine generated validation feedback information. The large language model adjusted the penetration coefficient of the scheme based on the feedback information. After adjustment, the schemes were re-validated. All three candidate schemes passed the validation and were determined as the target prediction schemes.

[0072] S5. Call upon the existing forecasting system to perform calculations: The system interface call module maps the three target forecast schemes to the interface parameters of the existing short- and medium-term hydrological forecasting system and generates a call request. After receiving the request, the existing system automatically obtains numerical weather forecast rainfall and reservoir measured flow data, calls the hydrological forecast calculation module to execute the forecast calculations of the three schemes respectively, obtains the 7-day inflow process line, flood peak flow, and peak occurrence time, and generates multi-scheme comparison and evaluation indicators (flood peak deviation, Nash efficiency coefficient).

[0073] S6. Intelligent interpretation of results and accumulation of experience: The large language model receives forecast results and comparison indicators of multiple scenarios, and combines them with historical flood cases in the knowledge base to generate natural language interpretation results: "The peak flow rates of the three scenarios in this forecast are 2500 m³ / s, 2480 m³ / s, and 2520 m³ / s, respectively, with the peak time being 14:00 on February 12, 2026. The Nash efficiency coefficients are all greater than 0.85, indicating high reliability of the forecast results; the highest forecast flood level is 104.5 m, which does not exceed the flood limit level of 105 m, indicating no flood risk; the second scenario is recommended, as its parameter configuration has the highest matching degree with similar historical floods." Meanwhile, the structured task description, three target forecast schemes, forecast results, and the second scheme finally adopted were processed in a structured manner, vectorized, and written back to the hydrological forecast knowledge base to update the vector retrieval index, thus completing the accumulation of forecast experience.

[0074] In this embodiment, forecasters only need to input their natural language requirements and do not need to configure any model parameters to obtain forecast results and professional interpretations, which greatly improves forecast efficiency and allows new staff to start forecasting work directly.

[0075] Example 5: This embodiment provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it executes the aforementioned human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model.

[0076] Example 6: This embodiment provides a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the aforementioned human-machine collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model.

[0077] It should be noted that the above examples are merely specific embodiments of the present invention, and the present invention is obviously not limited to the above embodiments, with many similar variations. All modifications that can be directly derived or conceived by those skilled in the art from the content disclosed in this invention should fall within the protection scope of this invention.

Claims

1. A large language model enhanced short-to-medium term hydrological forecast human-computer collaboration method, characterized in that, Includes the following steps: S1. Construct a hydrological forecast knowledge base for the target watershed, collect, structure, and vectorize historical forecast data for the target watershed, and establish a vector retrieval index. S2: Obtain the natural language forecast requirements input by forecasters, and after preprocessing, use a large language model to perform semantic parsing and feature extraction to generate a machine-processable structured forecast task description. S3, based on the structured forecast task description, retrieve similar historical forecast scenarios from the hydrological forecast knowledge base, generate candidate forecast schemes, and complete the automatic completion and adjustment of key parameters; S4. Input the candidate forecasting scheme into the business rule engine for rule verification. If it fails, the large language model combines the verification feedback information to iteratively correct it until a target forecasting scheme that meets the business rule constraints is generated. S5, convert the target forecast scheme into a call request that conforms to the existing short- and medium-term hydrological forecasting system interface specifications, call the short- and medium-term hydrological forecasting system to perform forecast calculations, and obtain forecast results and multi-scheme comparison evaluation indicators; S6. The large language model performs a comprehensive analysis of the forecast results and the multi-scheme comparative evaluation indicators to generate natural language interpretation results. At the same time, the structured information of the entire forecast process is written back to the hydrological forecast knowledge base to realize the accumulation and reuse of forecast experience.

2. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S1 specifically includes: S11, determine the target watershed and its control reservoirs, control sections and hydrological stations, and collect at least one original operational data from the following related to the target watershed: historical hydrological forecasting schemes, typical rain and flood cases, scheduling procedures, model parameter configurations, forecast evaluation results and consultation records. S12, perform data cleaning and structured processing on the original business data, and parse the information on time, spatial location, watershed name, reservoir name, forecast type, flood type, model type and evaluation index into structured fields; S13, Perform feature extraction and vectorization processing on text data in structured materials, including at least word segmentation, entity recognition, keyword extraction and vector encoding, to obtain text vectors that can be used for similarity retrieval; S14. Based on the obtained text vector, establish a vector retrieval index, associate and store the structured fields with the corresponding text vectors, and form a hydrological forecast knowledge base oriented towards the target watershed.

3. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S2 specifically includes: S21, receive natural language forecasting requests input by forecasters through a human-computer interaction interface. The natural language forecasting requests include at least one of the following: target reservoir or control section, forecast period, key indicators, and model preferences. S22, preprocess the natural language prediction requirements, including noise removal, unifying time and space representation, parsing technical terms, and obtaining standardized input text; S23, the standardized input text is input into the large language model, and the large language model performs semantic parsing and element extraction on the natural language forecast requirements to generate a structured forecast task description. The structured forecast task description is represented using a preset data structure and includes at least one target object identification field, forecast period configuration field, model combination field, input data source field, and output indicator field.

4. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S3 specifically includes: S31, Based on the structured forecast task description, extract the target watershed or reservoir identifier, forecast type, flood type, forecast period range, and model type preference search conditions; S32, using the large language model to access the hydrological forecast knowledge base and its vector retrieval index, and based on the retrieval conditions and text vector similarity, retrieve at least one historical forecast scheme and parameter combination that is similar to the current forecast scenario in terms of watershed, reservoir, flood type and model type; S33, based on the similar historical schemes, at least one candidate forecast scheme is generated by combining the large language model with the structured forecast task description. The candidate forecast scheme includes at least one of the following: forecast model type, model combination method, initial values ​​of key parameters, areal rainfall weight, and forecast period configuration. S34, complete and adjust the missing or incomplete parameters in the candidate forecast schemes so that each candidate forecast scheme has a complete parameter configuration that can be directly executed in the short- and medium-term hydrological forecasting system.

5. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S4 specifically includes: S41, The candidate forecast scheme is input into the business rule engine. The business rule engine obtains the corresponding rule configuration from the rule configuration library according to the identifier of the target reservoir or hydrological station. The rule configuration includes at least one or more of the following: forecast period boundary, model application scope, water balance constraint, watershed and reservoir safety constraint, and system resource constraint. S42, the rule execution engine performs a rule check on the candidate forecast scheme according to the rule configuration, and determines whether the forecast period setting exceeds the allowable range, whether the selected model type and combination method are in the allowable list, whether the key parameters fall within the allowable value range, and whether the water balance and flood control safety constraints are met. S43, if the candidate prediction scheme fails all rule checks, a verification feedback information containing the rule items that are not satisfied and their reasons is generated, and the verification feedback information is provided to the large language model. The large language model adjusts the candidate prediction scheme under the constraints of the verification feedback information to obtain a new candidate prediction scheme. S44. Repeat steps S41 to S43 until a target forecasting scheme that meets the preset business rules is generated.

6. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S5 specifically includes: S51, the target forecast scheme is parsed and mapped according to the interface specification of the short- and medium-term hydrological forecasting system to generate an interface call request containing model identifier, parameter group, forecast period, forecast step size and input data source configuration; S52, through the meteorological and hydrological data query module in the short-to-medium-term hydrological forecasting system, the rainfall, flow, water level and auxiliary factor data for the corresponding time period are automatically obtained based on the interface call request; S53, call the hydrological forecast calculation module of the existing short- and medium-term hydrological forecast system, complete the forecast calculation according to the target forecast scheme configuration, obtain the forecast result corresponding to the target forecast scheme, and generate multi-scheme comparison evaluation index when there are multiple candidate schemes.

7. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, Step S6 specifically includes: S61, the forecast results and the multi-scheme comparison evaluation indicators are sent to the big language model, and the big language model combines typical rain and flood cases and historical evaluation information in the hydrological forecast knowledge base to generate natural language interpretation results containing flood process characteristics descriptions, risk warnings and scheme optimization suggestions; S62, the description of this structured forecasting task, the target forecasting scheme, the forecasting results, the final adopted scheme and its evaluation indicators are stored in the hydrological forecasting knowledge base in a structured form, and the newly added text data are vectorized and indexed and updated to realize the continuous accumulation and iterative utilization of forecasting experience.

8. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, The structured forecasting task description is represented using a preset data structure, which includes at least one of the following fields: The target object identification field is used to identify the target reservoir, control section, or hydrological station; The forecast period configuration field is used to indicate the start and end times of the forecast and the forecast period tiers; The model combination field is used to indicate the lumped hydrological model, semi-distributed hydrological model, distributed hydrological model, or combination of different types of hydrological models used. The input data source field indicates at least one of the following data sources: real-time station rainfall, meteorological grid products, numerical weather forecast rainfall, and historical flow. Output indicator field, used to indicate at least one of the following output indicators: peak flow, peak time, flow or water level process line, and cumulative runoff.

9. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, The data in the hydrological forecast knowledge base carries at least one tag, including watershed tag, reservoir tag, forecast type tag, flood type tag, model type tag, and forecast effect tag. When the large language model performs the similar historical forecast scenario retrieval in step S3, it filters and sorts the candidate historical records based on the tags.

10. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model according to claim 1, characterized in that, The business rule engine includes a rule configuration library and a rule execution engine. The rule configuration library stores the upper limit of the forecast period for different watersheds or reservoirs, the allowed model types and model combinations, the allowed value range of key parameters, and the flood control safety boundary requirements. When the rule execution engine verifies the candidate forecast scheme, it queries the corresponding rule configuration based on the identifier of the target reservoir or hydrological station and performs rule checks.

11. The human-computer collaborative method for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 1, characterized in that, The natural language interpretation results generated by the large language model in step S6 include: a comparative explanation of the differences in peak flow, peak time and cumulative runoff of different forecast schemes; a risk warning that may exceed the design flow, flood limit water level or control water level; an explanation of the similarity between the current forecast process and typical historical rain and flood processes; and at least one of the following: a draft structured text report for forecast consultation or dispatch command.

12. A human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model, characterized in that: The system for performing the method according to any one of claims 1-11, the system comprising: The user interaction module is used to receive natural language forecast requests input by forecasters and to display forecast results and natural language interpretations. The large language model interaction module is used to communicate with the large language model, perform semantic parsing on the natural language forecasting requirements, generate a structured forecasting task description, and intelligently interpret the forecasting results after obtaining the forecasting results to generate natural language interpretation results. The hydrological forecasting knowledge base module is used to store historical forecasting schemes, typical rain and flood cases, scheduling procedures, model parameter configurations and forecasting evaluation results. It performs vectorization processing on text data, establishes a vector retrieval index, and provides knowledge retrieval services for the large language model interaction module. The business rules engine module is used to verify the candidate prediction schemes generated by the large language model interaction module according to preset business rules, and output the target prediction scheme or verification feedback information that meets the business rules. The system interface call module is used to convert the target forecast scheme into an interface call request that can be recognized by the short- and medium-term hydrological forecast system, call the meteorological and hydrological data query module and the hydrological forecast calculation module in the short- and medium-term hydrological forecast system, and obtain forecast results and multi-scheme comparison evaluation indicators. The data storage and log module is used to store the natural language forecasting requirements, the structured forecasting task description, the target forecasting scheme, the forecasting results and the final adopted scheme, and to write at least a portion of the information into the hydrological forecasting knowledge base module.

13. The human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 12, characterized in that, The large language model interaction module, hydrological forecast knowledge base module, business rule engine module, system interface call module, and data storage and log module are deployed in software form on a server or cloud platform, and the module functions are implemented by the processor executing computer program instructions.

14. The human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 12, characterized in that, The system is loosely coupled to the existing short- and medium-term hydrological forecasting system, and there is no need to modify the core architecture and model calculation module of the existing short- and medium-term hydrological forecasting system.

15. The human-machine collaborative system for short- and medium-term hydrological forecasting enhanced by a large language model as described in claim 12, characterized in that, The business rules engine module includes a rule configuration library and a rule execution engine. The rule configuration library stores differentiated business rules for different river basins and reservoirs, and the rule execution engine is used to match rules and perform scheme verification.

16. A computer device, characterized in that, It includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it performs a human-machine collaborative method for short- and medium-term hydrological forecasting with large language model enhancement as described in any one of claims 1-11.

17. A computer-readable storage medium having a computer program stored thereon, the computer program, when executed by a processor, implementing a human-computer collaborative method for short- and medium-term hydrological forecasting with large language model enhancement as described in any one of claims 1-11.

Images