Method and system for evaluating changes in hydrological regime based on large models

Abstract

The application discloses a water conservancy project influence hydrological regime change evaluation method and system based on a large model, obtains multi-source hydrological data of a target water conservancy project influence area, calculates index values of multiple hydrological regime evaluation indexes before and after water conservancy project construction, constructs a water conservancy project hydrology field prompt word template, inputs the multi-source hydrological data and the index values into a pre-training large language model, generates an initial evaluation result, constructs an RAG module containing a water conservancy project hydrology specification document, a historical evaluation case library and an expert knowledge base, and performs knowledge enhancement and verification on the initial evaluation result, calculates a hydrological regime change degree by using an improved change range method, performs self-adaptive optimization on RVA boundaries of each evaluation index by using hydrological seasonal characteristics output by the large language model, and fuses the change degree and an evaluation report generated by the large language model to output a comprehensive hydrological regime change evaluation report, so that the intelligent level, accuracy and interpretability of evaluation are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of water conservancy engineering technology, and in particular to a method and system for evaluating the impact of water conservancy projects on changes in hydrological conditions based on large-scale models. Background Technology

[0002] The construction and operation of water conservancy projects significantly alter the natural hydrological conditions of rivers, thereby affecting the structure and function of river ecosystems. Currently, the mainstream method for assessing hydrological changes is the IHA-RVA hydrological change index and range method proposed by Richter et al. However, existing IHA-RVA methods and their improvements rely entirely on numerical calculations, lacking semantic understanding and interpretation of hydrological change mechanisms. The assessment results are presented only in numerical form, failing to provide qualitative interpretations and analysis reports of hydrological changes. In the range method, the upper and lower target boundaries of each assessment index typically use fixed percentile thresholds, failing to fully consider the seasonal patterns and interannual fluctuations of hydrological processes. This leads to discrepancies in the applicability and accuracy of the assessment results across different seasons, and an inability to fully utilize multi-source heterogeneous hydrological information for comprehensive judgment, resulting in low assessment efficiency and accuracy. Summary of the Invention

[0003] The purpose of this invention is to solve the above-mentioned problems by designing a method and system for evaluating the impact of water conservancy projects on changes in hydrological conditions based on a large model.

[0004] To achieve the above objectives, the technical solution of the present invention further includes the following steps in the above-mentioned evaluation method for the impact of large-scale water conservancy projects on changes in hydrological conditions: Obtain multi-source hydrological data of the area affected by the target water conservancy project, and use the hydrological index calculation module to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. A prompt word template for the hydrological field of water conservancy projects is constructed. The multi-source hydrological data is preprocessed and dimensionally unified to obtain a structured hydrological data description. The structured hydrological data description, hydrological situation evaluation index values, and a preset hydrological situation evaluation task are input into a pre-trained large language model. The large language model generates an initial evaluation result of hydrological situation changes through a reasoning path that includes hydrological element identification, trend analysis, impact determination, and evaluation conclusion generation. The initial evaluation result includes at least the hydrological element identification result, the hydrological element change trend before and after the construction of the water conservancy project, the impact determination of the hydrological situation, and the natural language evaluation conclusion. A knowledge retrieval and enhancement RAG module in the field of hydrology is constructed. The RAG module is used to enhance and verify the knowledge of the initial evaluation results and generate enhanced evaluation results. Based on the enhanced evaluation results, the degree of change in hydrological situation after the construction of water conservancy projects is calculated using the improved RVA range method. The hydrological situation change rate is integrated with the initial evaluation results to output a comprehensive hydrological situation change evaluation report that includes both quantitative change rate and qualitative evaluation conclusions.

[0005] Furthermore, in the aforementioned method for evaluating the impact of water conservancy projects on changes in hydrological conditions based on a large model, the step of acquiring multi-source hydrological data of the target water conservancy project's impact area, and calculating hydrological condition evaluation index values ​​for the two periods before and after the construction of the water conservancy project using the hydrological index calculation module, includes: Obtain multi-source hydrological data of the area affected by the target water conservancy project, including at least long-sequence runoff data, water level data, and meteorological data before and after the construction of the water conservancy project; The multi-source hydrological data also includes one or more of satellite remote sensing data, radar rainfall data, and hydrological model simulation data; the multi-source hydrological data with different dimensions are subjected to time scale unification, spatial benchmark unification, normalization or standardization processing to convert numerical data and text data into structured hydrological data descriptions; the DeepSeek R1 large language model is used to perform data quality assessment and outlier labeling on different types of multi-source hydrological data. The data quality assessment adopts a hybrid assessment algorithm that combines semantic understanding and rule verification based on the large language model. The hydrological situation assessment indicators also include unstructured hydrological feature indicators automatically extracted from hydrological monitoring text reports based on the large language model. These unstructured hydrological feature indicators include at least descriptions of extreme hydrological events, descriptions of river morphology changes, and descriptions of eco-hydrological responses.

[0006] Furthermore, in the aforementioned method for evaluating the impact of water conservancy projects on hydrological situation changes based on a large model, the step of constructing a hydrological domain prompt word template for water conservancy projects involves inputting the multi-source hydrological data, hydrological situation evaluation index values, and a preset hydrological situation evaluation task into a pre-trained large language model. The large language model then generates initial evaluation results for hydrological situation changes, including: The pre-trained large language model is a DeepSeek R1 large language model that has been fine-tuned with corpus in the field of water conservancy engineering and hydrology. The training data used for fine-tuning includes at least hydrological calculation sheets for water conservancy projects, hydrological analysis reports, and hydrological code provisions; fine-tuning adopts low-rank adaptation (LoRA) and performs domain adaptation based on the DeepSeek R1 large language model.

[0007] Furthermore, in the aforementioned method for evaluating the impact of water conservancy projects on hydrological situation changes based on a large model, the step of constructing a hydrological domain prompt word template for water conservancy projects involves inputting the multi-source hydrological data, hydrological situation evaluation index values, and a preset hydrological situation evaluation task into a pre-trained large language model. The large language model then generates initial evaluation results for hydrological situation changes, including: The prompt word template should include at least three components: task instruction prompts, dataset statistics prompts, and dataset context prompts; The dataset contextual hints are based on runoff generation and confluence mechanisms and incorporate information on climate conditions, underlying surface features, and human activities.

[0008] Furthermore, in the aforementioned method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions, the construction of a hydrological knowledge retrieval and enhancement (RAG) module, and the use of the RAG module to enhance and validate the initial evaluation results to generate enhanced evaluation results, includes: The key evaluation statements in the initial evaluation results are converted into vector queries through a text embedding model, which is built based on the Sentence-BERT model; the RAG module includes at least water conservancy engineering hydrological specifications documents, a historical hydrological situation evaluation case library, and a hydrological expert knowledge base. In the knowledge base of the RAG module, a vector similarity retrieval algorithm is used to retrieve the reference document fragment with the highest semantic similarity to the vector query. The vector similarity retrieval algorithm uses cosine similarity or dot product similarity as a metric. The retrieved reference document fragments are used as context information and input together with the initial evaluation results into the large language model to generate intermediate evaluation results that have been verified by knowledge. Identify evaluation conclusions in intermediate evaluation results that conflict with hydrological specifications for water conservancy projects, mark the conflicting conclusions, and trigger a re-evaluation.

[0009] Furthermore, in the aforementioned evaluation method for the impact of water conservancy projects on hydrological situation changes based on large-scale models, the step of calculating the degree of hydrological situation change after the construction of the water conservancy project using the improved RVA range method based on the enhanced evaluation results includes: The long-series runoff data before the construction of water conservancy projects are used to perform hydrological seasonality pattern recognition by large language model. The seasonal component and interannual scale fluctuation component in the runoff sequence are extracted by the time series pattern recognition algorithm based on Transformer architecture, and the hydrological feature description of each season is output. Based on the description of hydrological seasonal characteristics, the reasonable fluctuation range of each evaluation index in different seasons is determined; Based on a reasonable fluctuation range, the upper and lower target boundaries of each evaluation indicator in the range of variation method are seasonally weighted and adjusted. The seasonal weighting adopts a weight allocation algorithm based on the runoff amplitude of each season.

[0010] Furthermore, in the aforementioned method for evaluating the impact of large-scale water conservancy projects on hydrological situation changes, the step of fusing the degree of hydrological situation change with the initial evaluation results to output a comprehensive hydrological situation change evaluation report containing both quantitative change and qualitative evaluation conclusions includes: The degree of change in hydrological situation is associated with and stored in conjunction with multi-source hydrological data and hydrological situation evaluation index values ​​to construct an incremental hydrological situation evaluation case library, which is used to update the historical hydrological situation evaluation case library of the RAG module.

[0011] Furthermore, in the evaluation system for the impact of water conservancy projects on changes in hydrological conditions based on a large model, the evaluation system includes the following modules: The hydrological data acquisition module is used to acquire multi-source hydrological data of the target water conservancy project's impact area. Based on the multi-source hydrological data, the hydrological index calculation module is used to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. The initial situation assessment module is used to construct prompt word templates for the hydrological field of water conservancy projects, preprocess and unify the dimensions of the multi-source hydrological data to obtain a structured hydrological data description; input the structured hydrological data description, hydrological situation assessment index values, and preset hydrological situation assessment tasks into a pre-trained large language model, and generate an initial assessment result of hydrological situation changes through the reasoning path of hydrological element identification, trend analysis, impact degree determination, and evaluation conclusion generation by the large language model; the initial assessment result includes at least the hydrological element identification result, the hydrological element change trend before and after the construction of the water conservancy project, the impact degree determination of the hydrological situation, and the natural language evaluation conclusion; The situation enhancement evaluation module is used to construct a knowledge retrieval and enhancement generation RAG module in the hydrological field. The RAG module is used to enhance and verify the initial evaluation results to generate enhanced evaluation results. The situation change perception module is used to calculate the degree of hydrological situation change after the construction of water conservancy projects based on the enhanced evaluation results and using the improved RVA range method. The evaluation report generation module is used to integrate the hydrological situation change degree with the initial evaluation results and output a comprehensive hydrological situation change evaluation report that includes quantitative change degree and qualitative evaluation conclusions.

[0012] Furthermore, in the large-model-based evaluation system for the impact of water conservancy projects on hydrological situation changes, the situation change perception module includes the following sub-modules: The extraction submodule is used to perform hydrological seasonality pattern recognition on long-sequence runoff data before the construction of water conservancy projects using a large language model. It uses a time series pattern recognition algorithm based on the Transformer architecture to extract the seasonal components and interannual scale fluctuation components in the runoff sequence and outputs hydrological feature descriptions for each season. The determination submodule is used to determine the reasonable fluctuation range of each evaluation index in different seasons based on the description of the hydrological seasonal characteristics. The adjustment submodule is used to seasonally weight the upper and lower target boundaries of each evaluation index in the range of variation method based on the reasonable fluctuation range. The seasonal weighting adopts a weight allocation algorithm based on the runoff amplitude of each season.

[0013] Furthermore, in the large-model-based evaluation system for the impact of water conservancy projects on changes in hydrological conditions, the evaluation report generation module includes the following sub-modules: The storage submodule is used to associate and store the degree of hydrological situation change with multi-source hydrological data and hydrological situation evaluation index values, and to build an incremental hydrological situation evaluation case library, which is used to update the historical hydrological situation evaluation case library of the RAG module.

[0014] Its beneficial effects lie in generating natural language evaluation reports through a large language model, realizing a dual-modal evaluation paradigm of quantitative calculation + qualitative interpretation, and making up for the shortcomings of traditional methods that rely solely on numerical calculation and lack semantic explanation of hydrological change mechanisms. A RAG knowledge enhancement module in the field of hydrology for water conservancy engineering was constructed, introducing normative documents, historical cases, and expert knowledge into the reasoning process of the large model. By vectorizing key statements in the initial evaluation results and performing semantic retrieval in the knowledge base, relevant document fragments retrieved are used as contextual information to jointly reason with the large model input, effectively solving the problem of insufficient knowledge in the water conservancy professional field of general large models, significantly suppressing domain knowledge illusion, and ensuring the standardization and reliability of evaluation conclusions. Utilizing the large language model's ability to identify hydrological seasonal characteristics, the index boundaries of the range of variation method were adaptively optimized. Traditional RVA methods use fixed percentile boundaries, which cannot reflect the seasonal differences in hydrological processes. This invention extracts the seasonal components of runoff using the time-series pattern recognition capabilities of a large language model based on the Transformer architecture. Based on this, the RVA boundaries of each indicator are seasonally weighted and adjusted, making the evaluation results more consistent with the natural laws of hydrological processes and improving the accuracy and efficiency of the evaluation. Attached Figure Description

[0015] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.

[0016] Figure 1 This is a schematic diagram of the first embodiment of the evaluation method for the impact of water conservancy projects on changes in hydrological conditions based on a large model, as described in this invention. Figure 2 This is a schematic diagram of the second embodiment of the evaluation method for the impact of water conservancy projects on changes in hydrological conditions based on a large model in this invention. Figure 3 This is a schematic diagram of the first embodiment of the evaluation system for the impact of water conservancy projects on changes in hydrological conditions based on a large model, as described in this invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0018] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms "one," "an," and "this" used herein may also include the plural forms. It should be further understood that the terminology used in this specification includes the presence of features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0019] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 As shown, a method for evaluating the impact of water conservancy projects on changes in hydrological conditions based on a large model is presented. This evaluation method includes the following steps: Step 101: Obtain multi-source hydrological data of the area affected by the target water conservancy project. Based on the multi-source hydrological data, use the hydrological index calculation module to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. Specifically, this embodiment acquires multi-source hydrological data of the area affected by the target water conservancy project. The multi-source hydrological data includes at least long-sequence runoff data, water level data, and meteorological data before and after the construction of the water conservancy project. Preferably, the multi-source hydrological data also includes one or more of satellite remote sensing data, radar rainfall data, and hydrological model simulation data. After acquiring the multi-source hydrological data, a large language model is used to perform data quality assessment and outlier labeling on different types of multi-source hydrological data. This data quality assessment employs a hybrid assessment algorithm combining semantic understanding and rule verification based on a large language model. On the one hand, the large language model is used to semantically judge the rationality of the data sequence; on the other hand, conventional data verification rules in the hydrological field are combined to detect outliers. The combined results of these two aspects provide a data quality assessment conclusion.

[0020] Based on multi-source hydrological data, the hydrological index calculation module calculates the values ​​of multiple hydrological situation evaluation indicators for two periods before and after the construction of the water conservancy project. The hydrological situation evaluation indicators include at least some or all of the indicators in the IHA indicator system. The IHA indicator system covers 32 indicators reflecting different dimensions of hydrological situation, such as monthly average flow, annual extreme flow, high flow pulse duration, low flow pulse duration, and flow change rate. Preferably, the hydrological situation evaluation indicators also include unstructured hydrological feature indicators automatically extracted from hydrological monitoring text reports based on a large language model. These unstructured hydrological feature indicators include at least descriptions of extreme hydrological events, descriptions of river channel morphology changes, and descriptions of eco-hydrological responses. Specifically, the hydrological monitoring text report is input into the large language model, and specially designed prompts guide the model to extract unstructured descriptive information related to changes in hydrological situation, quantifying it into feature indicators that can participate in subsequent evaluations.

[0021] Step 102: Construct a prompt word template for the hydrological field of water conservancy projects, preprocess and unify the dimensions of multi-source hydrological data to obtain a structured hydrological data description; input the structured hydrological data description, hydrological situation evaluation index values, and the preset hydrological situation evaluation task into a pre-trained large language model, and generate an initial evaluation result of hydrological situation changes through the reasoning path of hydrological element identification, trend analysis, impact degree determination, and evaluation conclusion generation by the large language model; the initial evaluation result includes at least the hydrological element identification result, the hydrological element change trend before and after the construction of the water conservancy project, the impact degree determination of the hydrological situation, and the natural language evaluation conclusion; Specifically, in this embodiment, multi-source hydrological data includes at least long-sequence runoff data, water level data, and meteorological data before and after the construction of water conservancy projects, and may also include one or more of satellite remote sensing data, radar rainfall data, and hydrological model simulation data. These data have different dimensions and spatiotemporal resolutions. Before inputting into the large language model, the multi-source hydrological data undergoes preprocessing and dimension unification: numerical data such as runoff, water level, precipitation, temperature, and evaporation are aligned according to a unified time scale and spatial benchmark, and then normalized or standardized; statistical features or index values ​​related to hydrological situation assessment are extracted from satellite remote sensing data, radar rainfall data, and hydrological model simulation data; and the aforementioned numerical features, index values, and text-based hydrological monitoring information are converted into structured hydrological data descriptions.

[0022] A prompt word template for the hydrological domain of water conservancy projects is constructed. Structured hydrological data descriptions, hydrological situation evaluation index values, and pre-set hydrological situation evaluation tasks are input into a pre-trained large language model. The large language model generates initial evaluation results of hydrological situation changes according to the reasoning path of "hydrological element identification → trend analysis → impact degree determination → evaluation conclusion generation." The initial evaluation results include at least: hydrological element identification results, hydrological element change trends before and after water conservancy project construction, impact degree determination of the hydrological situation, and evaluation conclusions in natural language form. Evaluation level labels or scores may also be included. The pre-trained large language model is selected from any one or more of DeepSeek R1, ChatGPT-4o, and Gemini 2; the pre-trained large language model is a large language model fine-tuned using hydrological corpus data from water conservancy projects. The training data used for fine-tuning includes at least hydrological calculation sheets, hydrological analysis reports, and hydrological standard provisions from water conservancy projects. Fine-tuning employs low-rank adaptation (LoRA) or prefix-tuning algorithms, performing domain adaptation based on a general large language model. The LoRA algorithm achieves efficient parameter fine-tuning by introducing a low-rank decomposition matrix into the weight matrix of the pre-trained model. It requires only a minimal number of newly added parameters to adapt the model to specialized tasks in the field of water conservancy and hydrology. The prefix tuning algorithm, on the other hand, achieves task adaptation by adding learnable, continuous cue vectors before the model's input layer.

[0023] The hydrological situation assessment task description includes: basic information about the water conservancy project, time scale information, spatial scope information, and key hydrological element types. The large language model adaptively generates a hydrological situation change analysis framework matching the current assessment task based on the task description. The prompt word template adopts a structured prompt word design method, and includes at least three components: task instruction prompts, dataset statistical prompts, and dataset context prompts. Task instruction prompts clarify the assessment objectives, output format, and assessment criteria to the large language model; dataset statistical prompts provide the model with a summary of the statistical characteristics of the hydrological data; and dataset context prompts explain the hydrogeographic background of the assessment area to the model. The dataset context prompts are based on runoff generation and confluence mechanisms and incorporate information on climate conditions, underlying surface characteristics, and human activities.

[0024] Preferably, the reasoning strategy employed in generating the initial evaluation results using the large language model is the CoT (Conceptual Chain) reasoning strategy. By designing step-by-step reasoning prompts, the large language model is guided to generate evaluation results according to the reasoning path of "hydrological element identification → trend analysis → impact degree determination → evaluation conclusion generation." Specifically, the model first identifies the key hydrological elements and their changing characteristics in the evaluation area; then it analyzes the changing trends of each hydrological element before and after the construction of the water conservancy project; next, it determines the impact degree of the hydrological situation based on the changing trends; and finally, it synthesizes the aforementioned analyses to generate a complete evaluation conclusion.

[0025] Step 103: Construct a knowledge retrieval and enhancement RAG module in the hydrological field, and use the RAG module to enhance and verify the initial evaluation results to generate enhanced evaluation results; Specifically, this embodiment constructs a hydrological knowledge retrieval and enhancement generation module, RAG. The RAG module includes at least the following: hydrological standards documents for water conservancy projects, a historical hydrological situation assessment case library, and a hydrological expert knowledge base. The hydrological standards documents for water conservancy projects include industry standard documents such as hydrological analysis and calculation specifications, environmental impact assessment specifications for water conservancy projects, and technical guidelines for determining ecological flow. The historical hydrological situation assessment case library contains past hydrological situation assessment cases and their evaluation results for various water conservancy projects. The hydrological expert knowledge base contains the evaluation experience, judgment rules, and professional knowledge accumulated by experts in the field of hydrology. The RAG module is used to enhance and verify the initial evaluation results, generating enhanced hydrological situation change evaluation results.

[0026] Knowledge enhancement and verification specifically includes the following sub-steps: First, key evaluation statements in the initial evaluation results are converted into vector queries using a text embedding model. The text embedding model is selected from BERT embedded models, Sentence-BERT, or domain-specific embedded models fine-tuned for the water conservancy field. Second, in the knowledge base of the RAG module, a vector similarity retrieval algorithm is used to retrieve the reference document fragment with the highest semantic similarity to the vector query. The vector similarity retrieval algorithm uses cosine similarity or dot product similarity as a metric. Third, the retrieved reference document fragments are used as contextual information and input together with the initial evaluation results into the large language model to generate intermediate evaluation results that have undergone knowledge verification. Fourth, evaluation conclusions in the intermediate evaluation results that conflict with water conservancy engineering hydrological specifications are identified, and conflicting conclusions are marked and a re-evaluation is triggered. Through the above RAG enhancement and verification mechanism, the potential knowledge illusion that the large language model may generate in the water conservancy professional field can be effectively suppressed, ensuring the standardization and reliability of the evaluation conclusions.

[0027] Step 104: Based on the enhanced evaluation results, calculate the degree of change in hydrological situation after the construction of the water conservancy project using the improved RVA range method; Specifically, in this embodiment, based on the enhanced hydrological situation change evaluation results, an improved Range of Variation (RVA) method is used to calculate the degree of hydrological situation change after the construction of water conservancy projects. The improved RVA method includes at least the following: adaptive optimization of the upper and lower target boundaries of each evaluation index based on the hydrological seasonality characteristics output by the large language model. In traditional RVA methods, the RVA boundaries of each hydrological index are fixed at the 25th and 75th percentile values ​​of the hydrological parameters before the impact. However, such fixed boundaries cannot reflect the seasonal differences in hydrological processes—the reasonable fluctuation range of the same index should be different in the wet and dry seasons. This invention utilizes the ability of the large language model to identify hydrological seasonality patterns to achieve adaptive optimization of the RVA boundaries.

[0028] The upper and lower target boundaries of each evaluation indicator are adaptively optimized based on the hydrological seasonal characteristics output by the large language model. This involves the following sub-steps: First, the large language model is used to perform hydrological seasonal pattern recognition on long-sequence runoff data prior to the construction of water conservancy projects. A time-series pattern recognition algorithm based on the Transformer architecture is used to extract the seasonal and interannual scale fluctuation components from the runoff sequence, outputting hydrological characteristic descriptions for each season. The Transformer architecture, through its self-attention mechanism, can effectively capture long-range dependencies and periodic patterns in the time series, exhibiting stronger pattern recognition capabilities compared to traditional seasonal decomposition methods. Second, based on the hydrological seasonal characteristic descriptions, the reasonable fluctuation ranges of each evaluation indicator in different seasons are determined. For example, for the monthly average flow indicator, the large language model can analyze historical data to derive reasonable flow fluctuation ranges for the wet and dry seasons. Third, based on the reasonable fluctuation ranges, the upper and lower target boundaries of each evaluation indicator in the range-of-variance (RVA) method are seasonally weighted. The seasonal weighting uses a weight allocation algorithm based on the runoff amplitude of each season; that is, seasons with larger runoff amplitudes are assigned higher weights, allowing the RVA boundary to dynamically reflect the hydrological change characteristics of different seasons.

[0029] Step 105: Integrate the hydrological situation change rate with the initial evaluation results to output a comprehensive hydrological situation change evaluation report that includes both quantitative change rate and qualitative evaluation conclusions.

[0030] Specifically, this embodiment integrates the hydrological situation change degree with the natural language evaluation report generated by the large language model, outputting a comprehensive hydrological situation change evaluation report that includes both quantitative change degree and qualitative evaluation conclusions. The integration process includes mapping the hydrological situation change degree to preset evaluation level labels, such as no change, slight change, moderate change, and severe change; then concatenating and integrating these evaluation level labels with the natural language evaluation report generated by the large language model to generate a comprehensive evaluation report containing evaluation level, change degree values, and textual interpretation. Through this integration method, the comprehensive evaluation report includes both objective quantitative evaluation results and qualitative interpretation and analysis based on the semantic understanding capabilities of the large language model, achieving a dual evaluation output driven by both data and knowledge.

[0031] Preferably, the method further includes associating and storing the comprehensive hydrological situation change assessment report with multi-source hydrological data and the index values ​​of hydrological situation assessment indicators to construct an incremental hydrological situation assessment case library, which is used to update the historical hydrological situation assessment case library of the RAG module. Through this incremental learning mechanism, as assessment cases are continuously accumulated, the knowledge base of the RAG module is continuously enriched and improved, and the knowledge enhancement effect of subsequent assessments will gradually improve.

[0032] Preferably, the method runs on a digital twin water conservancy system platform, and the comprehensive hydrological situation change assessment report is displayed and interacted with in a visual manner on the digital twin water conservancy system platform. Users can view the spatiotemporal distribution of hydrological situation changes, the degree of change of each indicator, the comprehensive evaluation level, and the text interpretation report generated by the large language model through the platform interface, and can perform interactive queries and in-depth analysis.

[0033] Its beneficial effects lie in generating natural language evaluation reports through a large language model, realizing a dual-modal evaluation paradigm of quantitative calculation + qualitative interpretation, and making up for the shortcomings of traditional methods that rely solely on numerical calculation and lack semantic explanation of hydrological change mechanisms. A RAG knowledge enhancement module in the field of hydrology for water conservancy engineering was constructed, introducing normative documents, historical cases, and expert knowledge into the reasoning process of the large model. By vectorizing key statements in the initial evaluation results and performing semantic retrieval in the knowledge base, relevant document fragments retrieved are used as contextual information to jointly reason with the large model input, effectively solving the problem of insufficient knowledge in the water conservancy professional field of general large models, significantly suppressing domain knowledge illusion, and ensuring the standardization and reliability of evaluation conclusions. Utilizing the large language model's ability to identify hydrological seasonal characteristics, the index boundaries of the range of variation method were adaptively optimized. Traditional RVA methods use fixed percentile boundaries, which cannot reflect the seasonal differences in hydrological processes. This invention extracts the seasonal components of runoff using the time-series pattern recognition capabilities of a large language model based on the Transformer architecture. Based on this, the RVA boundaries of each indicator are seasonally weighted and adjusted, making the evaluation results more consistent with the natural laws of hydrological processes and improving the accuracy and efficiency of the evaluation.

[0034] Please see Figure 2 In the evaluation method for the impact of water conservancy projects on hydrological situation changes based on large models, a knowledge retrieval and enhancement generation (RAG) module in the hydrological domain is constructed. This RAG module is used to enhance and validate the initial evaluation results, generating enhanced evaluation results through the following steps. Step 201: Convert the key evaluation statements in the initial evaluation results into vector queries using a text embedding model. The text embedding model is built based on the Sentence-BERT model. The RAG module should include at least the hydrological specifications document for water conservancy projects, a historical hydrological situation evaluation case library, and a hydrological expert knowledge base. Step 202: In the knowledge base of the RAG module, the vector similarity retrieval algorithm is used to retrieve the reference document fragment with the highest semantic similarity to the vector query. The vector similarity retrieval algorithm uses cosine similarity or dot product similarity as a measure. Step 203: Input the retrieved reference document fragments as context information and the initial evaluation results into the large language model to generate intermediate evaluation results that have been verified by knowledge. Step 204: Identify evaluation conclusions in the intermediate evaluation results that conflict with the hydrological specifications of water conservancy projects, mark the conflicting conclusions, and trigger a re-evaluation.

[0035] The above describes embodiments of the present invention's method for evaluating the impact of large-scale water conservancy projects on hydrological situation changes. Please refer to [link / reference]. Figure 3 In the large-scale model-based evaluation system for the impact of water conservancy projects on hydrological situation changes, the evaluation system includes the following modules: The hydrological data acquisition module is used to acquire multi-source hydrological data of the target water conservancy project's impact area. Based on the multi-source hydrological data, the hydrological index calculation module is used to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. The initial situation assessment module is used to construct prompt word templates for hydrological fields in water conservancy projects, preprocess and unify the dimensions of multi-source hydrological data to obtain structured hydrological data descriptions. The structured hydrological data descriptions, hydrological situation assessment index values, and preset hydrological situation assessment tasks are input into a pre-trained large language model. The large language model then generates initial assessment results of hydrological situation changes through a reasoning path that includes hydrological element identification, trend analysis, impact determination, and assessment conclusion generation. The initial assessment results include at least the hydrological element identification results, hydrological element change trends before and after water conservancy project construction, assessment of the impact of the hydrological situation, and natural language assessment conclusions. The situation enhancement assessment module is used to construct a knowledge retrieval and enhancement generation RAG module in the hydrological field. The RAG module is used to enhance and verify the initial assessment results and generate enhanced assessment results. The situation change perception module is used to calculate the degree of hydrological situation change after the construction of water conservancy projects based on the enhanced evaluation results and using the improved RVA range of change method. The evaluation report generation module is used to integrate the hydrological situation change rate with the initial evaluation results and output a comprehensive hydrological situation change evaluation report that includes both quantitative change rate and qualitative evaluation conclusions.

[0036] Acquire multi-source hydrological data for the area affected by the target water conservancy project. Specifically, collect daily runoff and water level data from hydrological stations for two periods: before the reservoir project's construction (1956-1995) and after construction (1996-2025), as well as daily precipitation, temperature, and evaporation data from meteorological stations within the basin. Simultaneously, acquire satellite remote sensing imagery of the area to aid in analyzing changes in the underlying surface of the basin. Integrate the above multi-source data into a data preprocessing module for data format standardization, missing value imputation, and outlier detection. In the data quality assessment stage, input each data sequence into a large language model. A hybrid assessment algorithm combining semantic understanding and rule validation is used to comprehensively evaluate data completeness, consistency, and rationality, and outliers are identified.

[0037] Based on preprocessed multi-source hydrological data, the IHA indicator system values ​​for the periods before and after the construction of the water conservancy project were calculated using a hydrological indicator calculation module. The IHA indicator system includes 32 hydrological indicators, covering five dimensions: monthly average flow, annual extreme flow, duration and frequency of high flow pulses, duration and frequency of low flow pulses, flow change rate, and number of reversals. Statistical values ​​for each indicator were calculated for both the pre-construction and post-construction periods. Simultaneously, historical monitoring text reports from hydrological monitoring stations were input into a large language model. Guided by prompts, the model extracted unstructured hydrological features such as extreme flood events, riverbed scouring and deposition changes, and fish migration responses, quantifying these features into auxiliary evaluation indicators.

[0038] A prompt word template for the hydrological domain of water conservancy projects was constructed. The prompt word template consists of three parts: a task instruction prompt that clearly informs the large language model of the reservoir name, geographical location, construction time, evaluation timescale, and spatial scope; a dataset statistical prompt that provides the model with statistical comparison data of various hydrological indicators before and after construction; and a dataset context prompt that, based on runoff generation and confluence mechanisms, describes the underlying surface characteristics of the evaluation area, such as climate type, multi-year average precipitation, watershed area, soil and vegetation cover, as well as upstream human activities. These prompt words, along with hydrological data and indicator values, are input into a DeepSeek R1 large language model fine-tuned using LoRA. During the inference process, a thought chain strategy is employed to guide the model through four inference steps: hydrological element identification, trend analysis, impact degree determination, and evaluation conclusion generation.

[0039] A hydrological RAG module is constructed. The module's knowledge base includes: normative documents such as the Ministry of Water Resources' hydrological analysis and calculation standards, reservoir environmental impact assessment standards, and ecological flow determination technical guidelines; past hydrological situation assessment case reports for this and similar watersheds; and evaluation experience rules and judgment criteria provided by hydrological experts. Key evaluation statements in the initial evaluation results generated in step S3 are converted into vectors using a Sentence-BERT model. Cosine similarity is then used for semantic retrieval in the knowledge base to obtain the most relevant normative clauses and case fragments. The retrieval results are then used as context, along with the initial evaluation results, and re-input into the large language model to generate knowledge-verified intermediate evaluation results. Subsequently, conclusions in the intermediate evaluation results that conflict with the normative documents are identified and marked, triggering a re-evaluation if necessary.

[0040] Based on the enhanced evaluation results, an improved RVA method is used to calculate the degree of hydrological situation variability. Traditional RVA methods use the 25th and 75th percentiles of each indicator before construction as fixed boundaries. In this invention, long-series runoff data before construction is first input into a large language model, and a time-series pattern recognition algorithm based on the Transformer architecture is used to extract the seasonal and interannual fluctuation components of runoff. The large language model then outputs descriptions of the hydrological characteristics for each season, such as the flood season in this watershed being from June to September, accounting for more than 70% of the annual runoff, and the dry season being from December to February of the following year, with runoff accounting for only about 5% of the annual runoff. Based on these seasonal characteristic descriptions, the reasonable fluctuation range of each IHA indicator in different seasons is determined.

[0041] The calculated hydrological situation change degree is mapped to a preset evaluation level label. The evaluation level label is then integrated with a natural language evaluation report generated by a large language model, which includes an analysis of the causes of change, an ecological impact assessment, and management recommendations. This results in a comprehensive evaluation report that includes the evaluation level, the change degree value, and a textual interpretation.

[0042] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely preferred examples and are not intended to limit the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended technical solutions and equivalents.

Claims

1. A method for evaluating the impact of water conservancy projects on changes in hydrological conditions based on a large-scale model, characterized in that, The evaluation method includes the following steps: Obtain multi-source hydrological data of the area affected by the target water conservancy project, and use the hydrological index calculation module to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. A prompt word template for the hydrological field of water conservancy projects is constructed. The multi-source hydrological data is preprocessed and dimensionally unified to obtain a structured hydrological data description. The structured hydrological data description, hydrological situation evaluation index values, and a preset hydrological situation evaluation task are input into a pre-trained large language model. The large language model generates an initial evaluation result of hydrological situation changes through a reasoning path that includes hydrological element identification, trend analysis, impact determination, and evaluation conclusion generation. The initial evaluation result includes at least the hydrological element identification result, the hydrological element change trend before and after the construction of the water conservancy project, the impact determination of the hydrological situation, and the natural language evaluation conclusion. A knowledge retrieval and enhancement RAG module in the field of hydrology is constructed. The RAG module is used to enhance and verify the knowledge of the initial evaluation results and generate enhanced evaluation results. Based on the enhanced evaluation results, the degree of change in hydrological situation after the construction of water conservancy projects is calculated using the improved RVA range method. The hydrological situation change rate is integrated with the initial evaluation results to output a comprehensive hydrological situation change evaluation report that includes both quantitative change rate and qualitative evaluation conclusions.

2. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, The process involves acquiring multi-source hydrological data of the target water conservancy project's impact area, and based on this data, using a hydrological index calculation module to calculate hydrological situation evaluation index values ​​for the two periods before and after the water conservancy project's construction. This includes: Obtain multi-source hydrological data of the area affected by the target water conservancy project, including at least long-sequence runoff data, water level data, and meteorological data before and after the construction of the water conservancy project; The multi-source hydrological data also includes one or more of satellite remote sensing data, radar rainfall data, and hydrological model simulation data; the multi-source hydrological data with different dimensions are subjected to time scale unification, spatial benchmark unification, normalization or standardization processing to convert numerical data and text data into structured hydrological data descriptions; the DeepSeek R1 large language model is used to perform data quality assessment and outlier labeling on different types of multi-source hydrological data. The data quality assessment adopts a hybrid assessment algorithm that combines semantic understanding and rule verification based on the large language model. The hydrological situation assessment indicators also include unstructured hydrological feature indicators automatically extracted from hydrological monitoring text reports based on the large language model. These unstructured hydrological feature indicators include at least descriptions of extreme hydrological events, descriptions of river morphology changes, and descriptions of eco-hydrological responses.

3. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, The structured hydrological data description, hydrological situation evaluation index values, and preset hydrological situation evaluation tasks are input into a pre-trained large language model. The large language model then generates an initial evaluation result of the hydrological situation change through a reasoning path based on hydrological element identification, trend analysis, impact determination, and evaluation conclusion generation. This includes: The pre-trained large language model is a DeepSeek R1 large language model that has been fine-tuned with corpus in the field of water conservancy engineering and hydrology. The training data used for fine-tuning includes at least hydrological calculation sheets for water conservancy projects, hydrological analysis reports, and hydrological code provisions; fine-tuning adopts low-rank adaptation (LoRA) and performs domain adaptation based on the DeepSeek R1 large language model.

4. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, The construction of a hydrological domain prompt word template for water conservancy projects involves inputting multi-source hydrological data, hydrological situation evaluation index values, and a preset hydrological situation evaluation task into a pre-trained large language model. The large language model then generates initial evaluation results of hydrological situation changes, including: The prompt word template should include at least three components: task instruction prompts, dataset statistics prompts, and dataset context prompts; The dataset contextual hints are based on runoff generation and confluence mechanisms and incorporate information on climate conditions, underlying surface features, and human activities.

5. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, The aforementioned construction of a knowledge retrieval and enhancement RAG module in the hydrological field is used to enhance and validate the initial evaluation results, generating enhanced evaluation results, including: The key evaluation statements in the initial evaluation results are converted into vector queries through a text embedding model, which is built based on the Sentence-BERT model; the RAG module includes at least water conservancy engineering hydrological specifications documents, a historical hydrological situation evaluation case library, and a hydrological expert knowledge base. In the knowledge base of the RAG module, a vector similarity retrieval algorithm is used to retrieve the reference document fragment with the highest semantic similarity to the vector query. The vector similarity retrieval algorithm uses cosine similarity or dot product similarity as a metric. The retrieved reference document fragments are used as context information and input together with the initial evaluation results into the large language model to generate intermediate evaluation results that have been verified by knowledge. Identify evaluation conclusions in intermediate evaluation results that conflict with hydrological specifications for water conservancy projects, mark the conflicting conclusions, and trigger a re-evaluation.

6. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, Based on the enhanced evaluation results, the improved RVA range method is used to calculate the degree of change in hydrological conditions after the construction of the water conservancy project, including: The long-series runoff data before the construction of water conservancy projects are used to perform hydrological seasonality pattern recognition by large language model. The seasonal component and interannual scale fluctuation component in the runoff sequence are extracted by the time series pattern recognition algorithm based on Transformer architecture, and the hydrological feature description of each season is output. Based on the description of hydrological seasonal characteristics, the reasonable fluctuation range of each evaluation index in different seasons is determined; Based on a reasonable fluctuation range, the upper and lower target boundaries of each evaluation indicator in the range of variation method are seasonally weighted and adjusted. The seasonal weighting adopts a weight allocation algorithm based on the runoff amplitude of each season.

7. The method for evaluating the impact of large-scale water conservancy projects on changes in hydrological conditions as described in claim 1, characterized in that, The process of integrating the hydrological situation change rate with the initial assessment results to output a comprehensive hydrological situation change assessment report containing both quantitative change rate and qualitative assessment conclusions includes: The degree of change in hydrological situation is associated with and stored in conjunction with multi-source hydrological data and hydrological situation evaluation index values ​​to construct an incremental hydrological situation evaluation case library, which is used to update the historical hydrological situation evaluation case library of the RAG module.

8. A large-scale model-based evaluation system for assessing the impact of water conservancy projects on changes in hydrological conditions, characterized in that... The evaluation system includes the following modules: The hydrological data acquisition module is used to acquire multi-source hydrological data of the target water conservancy project's impact area. Based on the multi-source hydrological data, the hydrological index calculation module is used to calculate the hydrological situation evaluation index values ​​for the two periods before and after the construction of the water conservancy project. The initial situation assessment module is used to construct prompt word templates for the hydrological field of water conservancy projects, preprocess and unify the dimensions of the multi-source hydrological data to obtain a structured hydrological data description; input the structured hydrological data description, hydrological situation assessment index values ​​and preset hydrological situation assessment tasks into a pre-trained large language model, and generate an initial assessment result of hydrological situation changes through the reasoning path of hydrological element identification, trend analysis, impact degree determination and assessment conclusion generation by the large language model; The initial evaluation results include at least the hydrological element identification results, the trend of hydrological element changes before and after the construction of the water conservancy project, the determination of the degree of impact of the hydrological situation, and the natural language evaluation conclusions. The situation enhancement evaluation module is used to construct a knowledge retrieval and enhancement generation RAG module in the hydrological field. The RAG module is used to enhance and verify the initial evaluation results to generate enhanced evaluation results. The situation change perception module is used to calculate the degree of hydrological situation change after the construction of water conservancy projects based on the enhanced evaluation results and using the improved RVA range method. The evaluation report generation module is used to integrate the hydrological situation change degree with the initial evaluation results and output a comprehensive hydrological situation change evaluation report that includes quantitative change degree and qualitative evaluation conclusions.

9. The evaluation system for the impact of large-scale water conservancy projects on hydrological situation changes as described in claim 8, characterized in that, The situation change perception module includes the following sub-modules: The extraction submodule is used to perform hydrological seasonality pattern recognition on long-sequence runoff data before the construction of water conservancy projects using a large language model. It uses a time series pattern recognition algorithm based on the Transformer architecture to extract the seasonal components and interannual scale fluctuation components in the runoff sequence and outputs hydrological feature descriptions for each season. The determination submodule is used to determine the reasonable fluctuation range of each evaluation index in different seasons based on the description of the hydrological seasonal characteristics. The adjustment submodule is used to seasonally weight the upper and lower target boundaries of each evaluation index in the range of variation method based on the reasonable fluctuation range. The seasonal weighting adopts a weight allocation algorithm based on the runoff amplitude of each season.

10. The evaluation system for the impact of large-scale water conservancy projects on hydrological situation changes as described in claim 8, characterized in that, The evaluation report generation module includes the following sub-modules: The storage submodule is used to associate and store the degree of hydrological situation change with multi-source hydrological data and hydrological situation evaluation index values, and to build an incremental hydrological situation evaluation case library, which is used to update the historical hydrological situation evaluation case library of the RAG module.

Images