Agricultural irrigation area water allocation method based on large model analysis

Abstract

The invention discloses an agricultural irrigation area water allocation method based on large model analysis, relates to the technical field of intelligent water conservancy and agricultural irrigation management, and solves the problems of low utilization efficiency caused by unreasonable water resource allocation in an irrigation area, lagging manual scheduling response and dependence on experience. Economic benefits and water supply safety are difficult to consider at the same time, and the problem that implementation is difficult due to poor interpretability of a scheduling scheme is solved. According to the method, a digital twinborn base fusing multi-modal data is constructed, the grid water demand is accurately calculated, a supply relation matrix is constructed by using a graph theory and a cost model, a multi-modal large model is trained to generate allocation rules and constraints, a space-time optimization model taking economic benefits and water shortage risks as targets is established, and a Pareto solution set is solved. And finally, a natural language decision suggestion containing multi-dimensional comparison and visual evidence is generated through a large model, and intelligent decision of water resource allocation in the irrigation area is realized.

Description

Technical Field

[0001] This invention belongs to the field of smart water conservancy and agricultural irrigation management technology, specifically a method for water allocation in agricultural irrigation areas based on large model analysis. Background Technology

[0002] Traditional water resource allocation methods mainly rely on human experience and static planning models, which are difficult to cope with complex dynamic factors such as weather, soil moisture, and crop water requirements. At the same time, they lack the ability to accurately quantify key factors such as water transmission costs and energy efficiency.

[0003] The existing technology has the following problems: low utilization efficiency due to unreasonable allocation of water resources in irrigation areas, slow response of manual scheduling and reliance on experience, decision-making dilemma of balancing economic benefits and water supply security, and implementation difficulties due to poor interpretability of scheduling schemes. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems existing in the prior art; to this end, the present invention proposes an agricultural irrigation water allocation method based on large model analysis to solve the above-mentioned technical problem.

[0005] The first aspect of this invention provides a method for water allocation in agricultural irrigation areas based on large-scale model analysis, comprising the following steps: S1: Divide agricultural irrigation areas into heterogeneous grid units and integrate their multimodal static and dynamic data to construct a digital twin foundation; S2: Based on the digital twin foundation, the Penman-Monteith model and water stress coefficient are driven to calculate the actual irrigation water demand of each grid and clarify the specific water demand grid unit; S3: Analyze the water conveyance paths from each water source to each water demand grid unit using a graph theory-based shortest path algorithm; analyze the energy consumption cost and efficiency loss of each water conveyance path using a hydraulic model and a cost quantification model; combine the water conveyance paths, energy consumption cost, and efficiency loss, and construct a supply relationship matrix between water sources and water demand grid units using an automated data integration method. S4: Construct a multimodal large model based on a vision-language architecture and complete pre-training and fine-tuning; after inputting real-time data from the digital twin base and the supply relationship matrix of the water source-demand grid unit, the multimodal large model outputs a set of priority allocation rules and key constraints for the water allocation method in agricultural irrigation areas. S5: Construct a spatiotemporal optimization model with dual optimization objectives. This model strictly follows the key constraints generated by the multimodal large model as hard boundary constraints and uses the priority allocation rule set as soft guiding constraints. A multi-objective evolutionary algorithm is used to solve the spatiotemporal optimization model to obtain the Pareto optimal solution set. S6: Multimodal large model deep fusion analysis of Pareto optimal solution set and real-time data in digital twin basis to generate natural language decision suggestions containing multi-dimensional comparative analysis and visual evidence to assist in the selection of final solution.

[0006] As a further aspect of the present invention: step S1 includes the following steps: The target agricultural irrigation area is divided into multiple uniform or non-uniform grid units based on crop type, soil properties, elevation, and canal system topology, and each grid unit is numbered; multimodal data, including static and dynamic data, are collected for each grid unit. The static data includes crop planting structure, soil water holding capacity, soil properties, crop root depth, water pump motor efficiency, upper limit of suitable moisture content, canal system topology, special data in the field of water conservancy and irrigation, and expert demonstration data. The soil properties include wilting coefficient and field capacity; the special data in the field of water conservancy and irrigation include water conservancy engineering drawings, canal system topology maps, water level process lines and structured descriptive text corresponding to the drawing reports, multi-period remote sensing images and their interpretation reports, and supply relationship matrices of water source-demand grid units containing information on multiple water source costs, water quality and engineering attributes; the expert demonstration data includes structured analysis reports provided by experts that provide multi-dimensional comparisons of agricultural irrigation water allocation schemes, visualization chart suggestions and contextualized recommendations, and historical agricultural irrigation water scheduling decision cases, each case containing multimodal data input and corresponding expert decision rule output; The dynamic data includes soil moisture, crop canopy coefficient, total head, meteorological data, real-time data near various water sources, and engineering operation data based on remote sensing inversion grid units; The meteorological data includes solar radiation, temperature, humidity, wind speed, precipitation, and soil heat flux; real-time data near each water source includes the real-time available water volume, water quality indicators, water intake costs, pumping head, and priority settings for different water sources for surface water, reclaimed water, and groundwater; and engineering operation data includes the dynamic water level depth of groundwater, the effluent pressure of the treatment plant, the real-time flow rate of channels or pipelines, the gate opening, the pumping station switch status, and energy consumption. The preprocessed multimodal static data and dynamic data are fused to construct a digital twin base.

[0007] As a further aspect of the present invention: step S2 includes the following steps: The saturated vapor pressure is calculated using temperature data, and the actual vapor pressure is derived by combining it with relative humidity data. The reference crop evapotranspiration is obtained using the standard analysis formula of the Penman-Monteith model based on solar radiation, soil heat flux, temperature, wind speed and vapor pressure difference. The real-time soil volumetric water content of grid cells is obtained by inversion using remote sensing technology. Combined with the upper limit of suitable crop water content and wilting coefficient, a water stress coefficient calculation model is established, and the water stress coefficient is obtained based on the real-time water content. The theoretical crop water requirement is obtained by multiplying the crop coefficient by the reference crop evapotranspiration, and then the actual crop evapotranspiration is obtained by correcting it with the water stress coefficient. At the same time, the change in soil water storage is calculated based on the changes in soil moisture content and crop root depth at different time periods. By solving the soil water balance equation, and taking into account the actual crop evapotranspiration, effective precipitation and soil water storage changes, the actual irrigation water demand of each grid unit is obtained. Whether a grid is a water-demand grid cell is determined by whether the actual irrigation water demand of the grid is greater than 0. If the actual irrigation water demand is greater than 0, then the grid is a water-demand grid cell.

[0008] As a further aspect of the present invention: step S3, which uses a graph theory-based shortest path algorithm to analyze the water transport path from each water source to each water-demand grid unit, includes the following steps: After identifying specific water demand grid units, their locations are determined by grid unit numbering. The entire water conveyance system is abstracted into a mathematical graph using network abstraction. Water sources, gates, pumping stations, and water demand grid units are treated as nodes. Channel segments or pipe segments connecting nodes are treated as variable edges, and their weights can be initialized to the physical length of the channel or pipe segment. The shortest path algorithm based on graph theory is used to efficiently find the shortest water conveyance path from any water source node to any water demand grid unit node. The shortest water conveyance paths found are then constructed into a graph database.

[0009] As a further aspect of the present invention: Step S3 involves analyzing the energy consumption cost and efficiency loss of each water conveyance path using a hydraulic model and a cost quantification model; combining the water conveyance path and the energy consumption cost and efficiency loss, an automated data integration method is used to construct a supply relationship matrix of water source-demand grid units, including the following steps: After constructing the graph database in the water conveyance path analysis, energy consumption and efficiency loss will be analyzed for each path based on the inherent characteristics of different types of water sources, using hydraulic and cost quantification models. First, energy consumption analysis will be performed, with the core being the pump station energy consumption calculation model. The graph database will be used to iterate through each determined water conveyance path to determine if it contains pump station facilities. If so, the operating parameters and real-time energy prices of the pump station will be extracted from the engineering database. Then, the energy cost per unit volume of water conveyed will be calculated using the pump energy consumption formula. Obtain energy consumption cost ;in, Indicates the density of water; Represents gravitational acceleration; Indicates total head; Indicates pump efficiency; Indicates motor efficiency; Indicates electricity price; Total head Different calculations are required depending on the type of water source. For surface water, it is the pumping head; for groundwater, it is the sum of the pumping head and the dynamic water level depth; for reclaimed water, it is the sum of the pumping head and the effluent pressure of the treatment plant. Secondly, an efficiency loss analysis is conducted, the core of which lies in the canal system water conveyance efficiency model. Based on the data information of the canal section and real-time meteorological data, a water conveyance efficiency coefficient with a value between 0 and 1 is assigned to it. ; By analyzing the formula: Obtain efficiency loss cost ;in, This represents the water conveyance efficiency coefficient; Indicates the price of the source water; Water source price It is not a single value; it needs to be dynamically determined based on the acquisition and treatment costs of different water sources. The price of surface water is a basic resource fee, the price of groundwater needs to include resource tax and ecological compensation costs, and the price of reclaimed water needs to cover the costs of advanced treatment and transportation. Finally, the energy consumption cost and efficiency loss cost are added together to obtain the total cost of transporting a unit of water from a specific type of water source to a specific water demand grid unit. After analyzing the comprehensive water transfer routes and their costs, an automated data integration method was used to construct a supply relationship matrix between water sources and water demand grid units. This matrix is ​​a three-dimensional structure, with row indices representing different water sources, column indices representing water demand grid units, and depth dimensions including key indicators such as total unit water transfer cost, route topology information, and water quality suitability index.

[0010] As a further aspect of the present invention: step S4 includes the following steps: A multimodal large model based on a vision-language architecture was constructed, and the model was pre-trained in a domain-adaptive manner using historical data integrated in a digital twin base. The training data included water conservancy engineering drawings, canal system topology maps, water level process lines, and structured descriptive text corresponding to the drawings and reports, multiple periods of remote sensing images and their interpretation reports, as well as a supply relationship matrix of water source-demand grid units containing information on the cost, water quality, and engineering attributes of multiple water sources. After pre-training, the large model was able to learn in depth the professional terminology, spatial relationships, physical mechanisms, characteristics of different water sources, and principles of coordinated utilization in the field of water conservancy and irrigation. After the multimodal large model is pre-trained, the multimodal large model is further trained using instruction fine-tuning and supervised learning methods. The training data comes from historical scheduling decision cases stored in the digital twin substrate. Each case contains multimodal data input and corresponding expert decision rule output. Through fine-tuning training, the multimodal large model can generate rule texts and constraints that conform to expert thinking and are applicable to multi-water source scenarios based on the input multimodal data. After the multimodal large model completes pre-training and fine-tuning, it is deployed in the decision loop of the digital twin base. In practical applications, the digital twin base is used as the data source, and the fused multimodal dynamic and static data and the supply relationship matrix of water source-demand grid units are input into the multimodal large model in real time. After analyzing these data, the multimodal large model outputs a set of priority allocation rules and key constraints to guide the allocation of water in agricultural irrigation areas.

[0011] As a further aspect of the present invention: step S5 includes the following steps: With three-dimensional matrix Based on this, a spatiotemporal optimization model with dual optimization objectives—maximizing the overall economic benefits of agricultural irrigation areas and minimizing total water shortage—is constructed. For water source indexing, For grid cell indexing, For time period indexing, that is, within the time period Inside, from the water source Distributed to grid cells Water volume; The first objective function aims to maximize economic benefits, expressed by the following function expression: ; To maximize economic benefits Indicates economic benefits; among which, Represents grid cells medium crops The unit benefit is expressed in yuan / kg; Output is expressed in kilograms; Indicates water source The unit cost of water, expressed in yuan / m³; This indicates the total amount of water diverted from this water source; This represents the energy consumption cost function for water transfer. Cs represents the inherent cost of water source s, where groundwater includes extraction energy costs, resource taxes, and depreciation and maintenance costs of well pump facilities; reclaimed water includes treatment costs and costs of pressurizing and maintaining the reclaimed water pipeline network; and surface water includes water rights fees and water source engineering maintenance costs. The second objective function aims to minimize the total water shortage, expressed as: To minimize the total water shortage, This represents the total water shortage; among which, Represents grid cells The actual irrigation water requirement, in cubic meters; This indicates the total amount of water actually supplied to this grid. The spatiotemporal optimization model strictly follows the key constraints generated by the multimodal large model as hard boundary constraints, and adjusts the priority allocation rule set by multiplying it by different weights according to the actual situation as soft guiding constraints. Among them, the hard boundary constraints include independent water supply capacity constraints, water conveyance capacity constraints, power grid load constraints, and crop-water source adaptability constraints based on water quality matching. The soft guiding constraints are transformed from the priority allocation rule set generated by the multimodal large model, including prioritizing the use of surface water, prioritizing the use of low-cost surface water, limiting the use of groundwater in over-extraction areas, and prioritizing the scheduling of gravity-flowable water sources during peak electricity consumption periods. Finally, based on the objectives of maximizing the overall economic benefits of agricultural irrigation areas and minimizing the total water shortage, the spatiotemporal optimization model integrates hard boundary constraints and soft guiding constraints, and uses a multi-objective evolutionary algorithm to obtain the Pareto optimal solution set for the water allocation method of agricultural irrigation areas.

[0012] As a further aspect of the present invention: step S6 includes the following steps: After the multimodal large model completes pre-training and fine-tuning, the structured Pareto optimal solution set and real-time multimodal data corresponding to the decision time, as well as the structured analysis report provided by experts, which compares these schemes from multiple dimensions, provides visualization charts and suggestions, and gives contextualized recommendations, are input into the multimodal large model for training. Among them, the key performance indicators of the Pareto optimal solution set need to clearly include the water consumption of each water source, the cost ratio of each water source, the intensity of groundwater extraction, and the utilization rate of reclaimed water for each scheme; the analysis report provided by experts needs to focus on comparing and demonstrating the water source structure, sustainability, and risks of different schemes. After the multimodal large model is trained, the Pareto optimal solution set obtained by solving the spatiotemporal optimization model and the real-time collected multimodal data are input into the multimodal large model and subjected to deep fusion analysis. The multimodal large model generates natural language decision suggestions that include multi-dimensional comparative analysis and visual evidence of water source utilization efficiency, economy and environmental sustainability, which assists in the selection of the final method for water allocation in agricultural irrigation areas.

[0013] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves a panoramic digital mapping of agricultural irrigation water resource systems through multimodal data fusion and digital twin infrastructure construction, enabling refined management of water resources. This invention achieves accurate simulation of crop water requirements and real-time response to water stress by combining mechanistic models with data-driven approaches. This invention solves the complex scheduling problem involving multiple water sources, multiple objectives, and multiple constraints by combining intelligent generation of scheduling rules from a multimodal large model with the collaborative solution of multi-objective optimization algorithms. This invention generates visualized and interpretable optimization schemes through a human-machine collaborative decision-making mechanism, thereby improving the intelligence level and decision-making efficiency of water resource allocation. Attached Figure Description

[0014] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0015] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0016] Please see Figure 1 The first aspect of this invention provides a method for allocating agricultural irrigation water based on large model analysis, comprising the following steps: S1: Divide agricultural irrigation areas into heterogeneous grid units and integrate their multimodal static and dynamic data to construct a digital twin foundation; S2: Based on the digital twin foundation, the Penman-Monteith model and water stress coefficient are driven to calculate the actual irrigation water demand of each grid and clarify the specific water demand grid unit; S3: Analyze the water conveyance paths from each water source to each water demand grid unit using a graph theory-based shortest path algorithm; analyze the energy consumption cost and efficiency loss of each water conveyance path using a hydraulic model and a cost quantification model; combine the water conveyance paths, energy consumption cost, and efficiency loss, and construct a supply relationship matrix between water sources and water demand grid units using an automated data integration method. S4: Construct a multimodal large model based on a vision-language architecture and complete pre-training and fine-tuning; after inputting real-time data from the digital twin base and the supply relationship matrix of the water source-demand grid unit, the multimodal large model outputs a set of priority allocation rules and key constraints for the water allocation method in agricultural irrigation areas. S5: Construct a spatiotemporal optimization model with dual optimization objectives. This model strictly follows the key constraints generated by the multimodal large model as hard boundary constraints and uses the priority allocation rule set as soft guiding constraints. A multi-objective evolutionary algorithm is used to solve the spatiotemporal optimization model to obtain the Pareto optimal solution set. S6: Multimodal large model deep fusion analysis of Pareto optimal solution set and real-time data in digital twin basis to generate natural language decision suggestions containing multi-dimensional comparative analysis and visual evidence to assist in the selection of final solution.

[0017] Specifically, the target agricultural irrigation area is divided into grids based on its internal heterogeneity, and real-time multimodal static and dynamic data are collected to construct a digital twin basis. The water stress coefficient is obtained through a computational model, and the reference crop evapotranspiration is obtained through the standard analysis formula of the Penman-Monteith model. Then, the soil water balance equation is solved to obtain the actual irrigation water demand for each grid unit. The agricultural irrigation area water conveyance system is abstracted as a mathematical graph, and the optimal water conveyance path from various surface water sources, groundwater sources, and reclaimed water sources to the water demand grid is obtained using a graph theory-based shortest path algorithm. The energy consumption cost and efficiency loss of the water conveyance process are analyzed using hydraulic and cost models. An automated data integration method is used to construct a water source-water demand grid unit supply relationship matrix containing multi-source cost and water quality information. A multimodal large-scale model based on a vision-language architecture was constructed and pre-trained and fine-tuned in the domain. After completing the domain pre-training and fine-tuning, the multimodal large-scale model can understand the professional terminology, spatial relationships, physical mechanisms, characteristics of different water sources, and principles of coordinated utilization in the field of water conservancy and irrigation, and output rule texts and constraints that conform to expert thinking. A dual-optimization objective spatiotemporal optimization model was constructed with the goals of maximizing economic benefits and minimizing total water shortage. This model integrates hard boundary constraints and soft guiding constraints, and uses a multi-objective evolutionary algorithm to solve it, obtaining the Pareto optimal solution set of the water allocation method for agricultural irrigation areas. The hard boundary constraints include independent water supply capacity constraints, water conveyance capacity constraints, power grid load constraints, and crop-water source adaptability constraints based on water quality matching. The soft guiding constraints are transformed from the priority allocation rule set generated by the multimodal large-scale model, including priority use of surface water, priority use of low-cost surface water, restriction of groundwater use in over-extraction areas, and priority scheduling of gravity-flowing water sources during peak electricity consumption periods. Finally, a multimodal large model is used to compare and visualize the Pareto optimal solution set from multiple dimensions and generate easy-to-understand natural language decision suggestions to assist in the selection of the final decision method.

[0018] In one embodiment of the present invention, step S1 includes the following steps: The target agricultural irrigation area is divided into multiple uniform or non-uniform grid units based on crop type, soil properties, elevation, and canal system topology, and each grid unit is numbered; multimodal data, including static and dynamic data, are collected for each grid unit. The static data includes crop planting structure, soil water holding capacity, soil properties, crop root depth, water pump motor efficiency, upper limit of suitable moisture content, canal system topology, special data in the field of water conservancy and irrigation, and expert demonstration data. The soil properties include wilting coefficient and field capacity; the special data in the field of water conservancy and irrigation include water conservancy engineering drawings, canal system topology maps, water level process lines and structured descriptive text corresponding to the drawing reports, multi-period remote sensing images and their interpretation reports, and supply relationship matrices of water source-demand grid units containing information on multiple water source costs, water quality and engineering attributes; the expert demonstration data includes structured analysis reports provided by experts that provide multi-dimensional comparisons of agricultural irrigation water allocation schemes, visualization chart suggestions and contextualized recommendations, and historical agricultural irrigation water scheduling decision cases, each case containing multimodal data input and corresponding expert decision rule output; The dynamic data includes soil moisture, crop canopy coefficient, total head, meteorological data, real-time data near various water sources, and engineering operation data based on remote sensing inversion grid units; The meteorological data includes solar radiation, temperature, humidity, wind speed, precipitation, and soil heat flux; real-time data near each water source includes the real-time available water volume, water quality indicators, water intake costs, pumping head, and priority settings for different water sources for surface water, reclaimed water, and groundwater; and engineering operation data includes the dynamic water level depth of groundwater, the effluent pressure of the treatment plant, the real-time flow rate of channels or pipelines, the gate opening, the pumping station switch status, and energy consumption. The preprocessed multimodal static data and dynamic data are fused to construct a digital twin base.

[0019] Specifically, based on the spatial distribution of crop types, spatial variability of soil properties, spatial heterogeneity of the elevation digital terrain model, and canal system topology within the target agricultural irrigation area, an adaptive grid partitioning algorithm is used to divide the irrigation area into multiple uniform or non-uniform heterogeneous grid units. Multimodal static and dynamic data are collected for each grid unit. Static data includes crop planting structure, soil water holding capacity, soil properties, crop root depth, pump motor efficiency, upper limit of suitable moisture content, canal system topology, special data in the field of water conservancy and irrigation, and expert demonstration data. Crop planting structure data comes from agricultural department archives and field surveys, specifically including crop types, varieties, and spatial distribution. Soil properties data comes from soil surveys and laboratory analyses conducted by the natural resources department, specifically including field capacity, wilting coefficient, and saturated hydraulic conductivity. Canal system topology data comes from the water conservancy department and engineering surveying, specifically including the geometric connections and properties of channels, pipelines, gates, and pumping stations. Engineering attribute data near various water sources comes from the water conservancy department, specifically including the location and elevation of surface water intakes, and the location and elevation of groundwater wells. The static water level, location of the reclaimed water treatment plant, and outflow pressure are all included. Specific data for the water conservancy and irrigation field comes from drawings by water conservancy departments and design institutes, as well as remote sensing satellite data. This includes historical water conservancy engineering drawings, canal system topology maps, water level process curves, structured descriptive text corresponding to the drawings, multi-period remote sensing images and their interpretation reports, and supply relationship matrices of water source-demand grid units containing information on multiple water source costs, water quality, and engineering attributes. Expert demonstration data comes from domain experts, specifically including structured analysis reports provided by experts that offer multi-dimensional comparisons of agricultural irrigation water allocation schemes, visualization chart suggestions, and contextualized recommendations, as well as historical agricultural irrigation water allocation decision-making cases. Each case includes multimodal data input and corresponding expert decision-making rule outputs. Dynamic data includes soil moisture, crop canopy coefficient, meteorological data, real-time data near each water source, and engineering operation data obtained from remote sensing image inversion. Meteorological data comes from meteorological stations and includes solar radiation, air temperature, humidity, wind speed, precipitation, and soil heat flux. Engineering operation data is acquired in real time through the SCADA system and each water source monitoring station and includes groundwater dynamic water level depth, treatment plant effluent pressure, real-time flow rate of channels or pipelines, gate opening, pump station on / off status, and energy consumption. Real-time data near each water source includes real-time surface water availability from river flow stations, real-time reclaimed water availability estimated from current dynamic water level and pump well output capacity, real-time groundwater availability from treatment plant flow meters, water quality indicators from automatic water quality monitoring instruments, water intake costs from water resource management departments, pump head, and priority settings for different water sources.

[0020] The multimodal static and dynamic data are spatiotemporally aligned, standardized, and fused: static attribute data is associated with dynamic monitoring data using the unique geocoding of grid cells; kriging interpolation and data assimilation techniques are used to interpolate point sensor data into area grid data, which is then fused with area data retrieved from remote sensing; finally, a unified spatiotemporal benchmark, multidimensional, and high-fidelity digital twin is constructed to provide data support for subsequent calculations.

[0021] In one embodiment of the present invention, step S2 includes the following steps: Using Tetens' formula: Obtain saturated water vapor pressure ;in, Indicates temperature; By analyzing the formula: Obtain the actual water vapor pressure ;in, Indicates relative humidity. Indicates saturated water vapor pressure; Using the standard analysis formula of the Penman-Monteith model: ; Obtain reference crop evapotranspiration ;in, Indicates solar radiation; Indicates soil heat flux; Indicates temperature; Indicates wind speed; Indicates saturated water vapor pressure; Indicates the actual water vapor pressure; The water stress coefficient was obtained through a computational model. ,when hour, =1; when hour, =0; when hour, Where θ represents the real-time soil volumetric water content obtained from remote sensing inversion; Indicates the upper limit of suitable moisture content; Indicates the wilting coefficient; By analyzing the formula: Theoretical crop water requirements ;in, Represents the crop coefficient; Represents reference crop evapotranspiration; By analyzing the formula: Obtain actual crop evapotranspiration ;in, Indicates the water stress coefficient; This indicates the theoretical water requirement for crops; By analyzing the formula: Obtain soil water storage change ;in, Indicates soil moisture content at different time periods; Indicates the depth of the main root system of a crop, in meters; By solving the soil water balance equation: Obtain the actual irrigation water demand The unit is cubic meters; among which, This represents the actual crop evapotranspiration. Indicates effective precipitation; Indicates the change in soil water storage; This represents the grid area, in square meters. Indicates the grid cell number; By determining the actual irrigation water demand of this grid Whether it is a water-demanding grid cell is determined by whether it is greater than 0. If so, then the grid is a water-demand grid cell.

[0022] Specifically, using the Tetens formula Calculate the saturated vapor pressure, where T represents the air temperature, in units of... The physical relationship between air temperature and saturated vapor pressure is described by an exponential relationship, reflecting the characteristic that the air's ability to hold water vapor increases with rising temperature. By inputting real-time air temperature data from the grid cells, saturated vapor pressure values ​​with an accuracy of ±0.05 kPa can be obtained, in kPa, providing basic parameters for subsequent calculations.

[0023] Based on formula Calculate the actual water vapor pressure, where RH represents relative humidity in %; this calculation is based on the definition of relative humidity, where the actual water vapor pressure is equal to the product of the saturated water vapor pressure and the relative humidity. This calculation can accurately characterize the actual water vapor content in the air and is a key parameter for calculating the driving force of evapotranspiration.

[0024] Using the standard formula of the Penman-Monteith model Calculate the reference crop evapotranspiration; the standard formula for this model combines the energy balance method and the aerodynamic method, in which... This represents the net radiation term, which provides the energy required for evaporation. Representing aerodynamics, it reflects the evapotranspiration effect caused by airflow; by inputting parameters such as solar radiation Rn, soil heat flux G, air temperature T, and wind speed u2, a reference crop evapotranspiration with an accuracy of ±0.5 mm / d can be obtained.

[0025] The water stress coefficient was obtained by calculating the piecewise function model. :when hour, ;when hour, ;when hour, This computational model is based on the soil-plant-atmosphere continuum theory and reflects the degree to which soil moisture conditions inhibit crop evapotranspiration. By inputting real-time volumetric soil moisture content θ retrieved from remote sensing and crop moisture parameters, the degree of water stress can be quantified.

[0026] Use formula Calculate the theoretical crop water requirement, where Kc represents the crop coefficient; convert the reference evapotranspiration into the theoretical water requirement of a specific crop using the crop coefficient to obtain the crop water requirement under ideal water conditions, in mm / d.

[0027] Based on formula Calculate the actual crop evapotranspiration; correct the theoretical water requirement by the water stress coefficient to reflect the impact of actual water conditions on crop evapotranspiration, and obtain a crop evapotranspiration that is closer to the actual situation, in mm / d.

[0028] Using formula Calculate the change in soil water storage, where Zroot represents the depth of the main crop root system in meters; based on the concept of soil reservoir, the net change in soil moisture in the root layer during the calculation period can be obtained, with positive values ​​representing soil water supply and negative values ​​representing soil water storage.

[0029] Through soil water balance equation The actual irrigation water requirement can be obtained by solving the problem. Where Pe represents effective precipitation; Am represents the grid area in square meters; this calculation is based on the principle of soil water balance, comprehensively considering evapotranspiration, precipitation recharge, and soil moisture changes, ultimately obtaining the actual irrigation water demand of the grid in cubic meters. Then, determine the actual irrigation water demand. Is it greater than 0? If the value is zero, then the grid cell is determined to be a water-demanding grid cell, and its value is zero. This is the actual irrigation water demand of the water demand grid unit; In one embodiment of the present invention, step S3, which uses a graph-based shortest path algorithm to analyze the water transport path from each water source to each water-demand grid unit, includes the following steps: After identifying specific water demand grid units, their locations are determined by grid unit numbering. The entire water conveyance system is abstracted into a mathematical graph using network abstraction. Water sources, gates, pumping stations, and water demand grid units are treated as nodes. Channel segments or pipe segments connecting nodes are treated as variable edges, and their weights can be initialized to the physical length of the channel or pipe segment. The shortest path algorithm based on graph theory is used to efficiently find the shortest water conveyance path from any water source node to any water demand grid unit node. The shortest water conveyance paths found are then constructed into a graph database.

[0030] Specifically, after identifying the specific water demand grid units, their locations are determined by grid unit numbers, forming a set of water demand grid units. These water demand grid units are the endpoints for water conveyance path calculation. Using network abstraction, the entire irrigation district water conveyance system is abstracted into a mathematical graph model G=(V,E). The vertex set V represents the nodes in the system, including water sources, gates, pumping stations, and water intake points of the water demand grid units; the edge set E represents the canal or pipe segments connecting these nodes, and each edge is assigned a weight value w(e). In the initial stage of path search, this weight is usually initialized to the physical length of the canal or pipe segment, thus intuitively representing the spatial distance cost of water conveyance. A graph-based shortest path algorithm is used to calculate the shortest physical path from each water source node to each water demand grid unit node. This algorithm can reliably find the path with the minimum cumulative weight. All calculated shortest paths, including path sequences (which nodes and edges are passed through) and path attributes (total length), are constructed into a graph database. This step uses graph theory to transform the complex physical water conveyance network into a computable mathematical model. Through the shortest path algorithm, the theoretically shortest water conveyance route can be systematically found, providing a basic solution for reducing water conveyance energy consumption and time.

[0031] In one embodiment of the present invention, step S3 involves analyzing the energy consumption cost and efficiency loss of each water conveyance path using a hydraulic model and a cost quantification model; combining the water conveyance path and the energy consumption cost and efficiency loss, an automated data integration method is used to construct a supply relationship matrix of water source-demand grid units, including the following steps: After constructing the graph database in the water conveyance path analysis, energy consumption and efficiency loss will be analyzed for each path based on the inherent characteristics of different types of water sources, using hydraulic and cost quantification models. First, energy consumption analysis will be performed, with the core being the pump station energy consumption calculation model. The graph database will be used to iterate through each determined water conveyance path to determine if it contains pump station facilities. If so, the operating parameters and real-time energy prices of the pump station will be extracted from the engineering database. Then, the energy cost per unit volume of water conveyed will be calculated using the pump energy consumption formula. Obtain energy consumption cost ;in, This indicates the density of water. Represents gravitational acceleration. Indicates total head. Indicates pump efficiency. Indicates motor efficiency. Indicates electricity price; Total head Different calculations are required depending on the type of water source. For surface water, it is the pumping head; for groundwater, it is the sum of the pumping head and the dynamic water level depth; for reclaimed water, it is the sum of the pumping head and the effluent pressure of the treatment plant. Secondly, an efficiency loss analysis is conducted, the core of which lies in the canal system water conveyance efficiency model. Based on the data information of the canal section and real-time meteorological data, a water conveyance efficiency coefficient with a value between 0 and 1 is assigned to it. ; By analyzing the formula: Obtain efficiency loss cost ;in, This represents the water conveyance efficiency coefficient; Indicates the price of the source water; Water source price It is not a single value; it needs to be dynamically determined based on the acquisition and treatment costs of different water sources. The price of surface water is a basic resource fee, the price of groundwater needs to include resource tax and ecological compensation costs, and the price of reclaimed water needs to cover the costs of advanced treatment and transportation. Finally, the energy consumption cost and efficiency loss cost are added together to obtain the total cost of transporting a unit of water from a specific type of water source to a specific water demand grid unit. After analyzing the comprehensive water transfer routes and their costs, an automated data integration method was used to construct a supply relationship matrix between water sources and water demand grid units. This matrix is ​​a three-dimensional structure, with row indices representing different water sources, column indices representing water demand grid units, and depth dimensions including key indicators such as total unit water transfer cost, route topology information, and water quality suitability index.

[0032] Specifically, after constructing the graph database of water conveyance routes, a refined cost analysis process based on hydraulic and cost quantification models begins. This aims to quantify the energy consumption and efficiency loss costs of each determined shortest water conveyance route. First, energy consumption analysis is performed, with the core being the pump station energy consumption calculation model. Each route is traversed using the graph database, checking if it contains pump station nodes. If a pump station is detected, its real-time operating parameters and the corresponding real-time energy market price are automatically extracted from the pre-built engineering database. Among them, the real-time operating parameters include the total head. Pump efficiency Motor efficiency For groundwater pathways, total head The dynamic water level depth needs to be factored in; for reclaimed water paths, it's necessary to consider whether the treatment plant's outlet pressure can be utilized to reduce the pumping head. Subsequently, a pump energy consumption formula derived from fluid mechanics is applied for calculation. This formula is based on the fundamental physical principle that a pump needs to overcome gravity to lift a unit volume of water. It calculates the ideal energy ρ*g*H required to lift a unit mass of water to a certain height, then divides by the combined efficiency of the pump and motor. The actual electrical energy consumed is obtained, and then multiplied by the electricity price to get the energy cost. Its operation involves a direct mathematical calculation; the inputs are the pump operating parameters and real-time electricity price, and the output is the energy cost per unit volume of water. This directly transforms the pump station's energy consumption into an economic indicator, providing a core basis for comparing the economics of different water conveyance routes. Secondly, efficiency loss analysis is performed, with the core being the canal system water conveyance efficiency model. Based on canal segment attribute data from the map database and combined with real-time meteorological data—including lining type, service life, and maintenance status, and temperature and wind speed—evaporation losses are estimated. Using a pre-set canal system efficiency comparison table or empirical formula, each canal segment in the route is assigned a water conveyance efficiency coefficient between 0 and 1. The overall efficiency coefficient of the entire route is the product of the coefficients of each canal segment; the principle of the efficiency loss cost formula is to consider the water lost due to leakage and evaporation during transportation as an economic cost, and its application involves considering the overall efficiency coefficient of the route. Substituting the inherent cost Cs of a unit volume of water from a specific water source s into the formula, the intangible physical loss is quantified into a concrete economic cost, clarifying the value loss of water resources during transportation. Finally, the calculated energy consumption cost and efficiency loss cost are superimposed to obtain the total cost C{s,g} of transporting a unit volume of water from a specific water source s to a specific water-demand grid unit g. After completing the cost analysis of all water source-demand grid unit path pairs, automated data integration is used to integrate the path information and its corresponding total cost C{s,g} of unit volume water transportation into a structured, machine-readable supply relationship matrix M of water source-demand grid units. In this matrix, the row index represents the water source, the column index represents the water-demand grid unit, and the matrix element M[s][g]=C{s,g} is the corresponding total cost of unit water transportation. The depth dimension includes key indicators such as total cost of unit water transportation, path topology information, and water quality suitability index.

[0033] In one embodiment of the present invention, step S4 includes the following steps: A multimodal large model based on a vision-language architecture was constructed, and the model was pre-trained in a domain-adaptive manner using historical data integrated in a digital twin base. The training data included water conservancy engineering drawings, canal system topology maps, water level process lines, and structured descriptive text corresponding to the drawings and reports, multiple periods of remote sensing images and their interpretation reports, as well as a supply relationship matrix of water source-demand grid units containing information on the cost, water quality, and engineering attributes of multiple water sources. After pre-training, the large model was able to learn in depth the professional terminology, spatial relationships, physical mechanisms, characteristics of different water sources, and principles of coordinated utilization in the field of water conservancy and irrigation. After the multimodal large model is pre-trained, the multimodal large model is further trained using instruction fine-tuning and supervised learning methods. The training data comes from historical scheduling decision cases stored in the digital twin substrate. Each case contains multimodal data input and corresponding expert decision rule output. Through fine-tuning training, the multimodal large model can generate rule texts and constraints that conform to expert thinking and are applicable to multi-water source scenarios based on the input multimodal data. After the multimodal large model completes pre-training and fine-tuning, it is deployed in the decision loop of the digital twin base. In practical applications, the digital twin base is used as the data source, and the fused multimodal dynamic and static data and the supply relationship matrix of water source-demand grid units are input into the multimodal large model in real time. After analyzing these data, the multimodal large model outputs a set of priority allocation rules and key constraints to guide the water allocation method in agricultural irrigation areas.

[0034] Specifically, a general vision-language model is selected as the base model to create a multimodal large model. Historical data integrated in the digital twin base is used for domain-adaptive pre-training. The training data includes, but is not limited to, water conservancy engineering design drawings, canal system topology diagrams, historical water level and flow process lines, and structured descriptive text corresponding to all these drawings and reports, multi-period remote sensing images and their interpretation reports, and supply relationship matrices of water source-demand grid units containing information on multiple water source costs, water quality, and engineering attributes. Through pre-training tasks of masked language modeling and image-text contrastive learning, the model learns and embeds professional terms in the field of water conservancy and irrigation, such as head, gate opening, and critical periods of crop water demand, complex spatial relationships such as canal connectivity and the relative positions of pumping stations and irrigation areas, and basic physical mechanisms such as water flow and crop water consumption patterns, thereby gaining a deep understanding and representation ability of multimodal data in the field of agricultural irrigation. After pre-training, the model was fine-tuned using a combination of instruction fine-tuning and supervised learning strategies. The training data came from historical scheduling decision cases stored in the digital twin substrate. These cases were constructed in the form of instruction-output pairs: the instruction part was a multimodal description of a specific scheduling scenario, such as the current soil moisture distribution map, reservoir water level process line, available water sources, and cost matrix as inputs; the output part was the scheduling rules and constraints formulated by experts under that scenario, such as prioritizing the use of low-cost surface water, limiting groundwater extraction in over-extraction areas, prioritizing water use in high-yield crop areas, limiting the operating time of high-lift pumping stations, and ensuring that the total water diversion volume does not exceed the channel design flow. After multiple training iterations, the multimodal large model learned and imitated the decision-making thinking patterns of experts, ultimately gaining the ability to generate rule texts and mathematical constraints that conform to irrigation scheduling logic and are applicable to multi-water source scenarios based on real-time multimodal data inputs. Ultimately, the fully trained multimodal large model is deployed on the decision support system based on the digital twin, forming a closed loop. During actual operation, the digital twin base serves as the data source, continuously feeding the model with real-time fused multimodal data, including the latest meteorological information, remote sensing inversion of soil moisture, engineering operation status, real-time available water volume, water quality and water intake costs of each water source, and the calculated supply-demand grid unit supply relationship cost matrix. After comprehensively analyzing this information, the multimodal large model outputs structured decision elements, namely a set of priority allocation rules and key constraints to guide the water allocation method in agricultural irrigation areas, thereby providing intelligent and adaptive high-level guidance for subsequent optimization solutions.

[0035] In one embodiment of the present invention, step S5 includes the following steps: With three-dimensional matrix Based on this, a spatiotemporal optimization model with dual optimization objectives—maximizing the overall economic benefits of agricultural irrigation areas and minimizing total water shortage—is constructed. For water source indexing, For grid cell indexing, For time period indexing, that is, within the time period Inside, from the water source Distributed to grid cells Water volume; The first objective function aims to maximize economic benefits: By analyzing the formula: Obtain the energy consumption cost of water transportation ,in, This indicates the total amount of water transported, i.e. The unit is cubic meters; Indicates the density of water; Represents gravitational acceleration; Indicates total head; Indicates pump efficiency; Indicates motor efficiency; Indicates electricity price; Total head Different calculations are required depending on the type of water source. For surface water, it is the pumping head; for groundwater, it is the sum of the pumping head and the dynamic water level depth; for reclaimed water, it is the sum of the pumping head and the effluent pressure of the treatment plant. Through function expressions: ; To maximize economic benefits Indicates economic benefits; among which, Represents grid cells medium crops The unit benefit is expressed in yuan / kg; Output is expressed in kilograms; Indicates water source The unit cost of water, expressed in yuan / m³; This indicates the total amount of water diverted from this water source, i.e. ; This represents the energy consumption cost function for water transfer. Cs represents the inherent cost of water source s, where groundwater includes extraction energy costs, resource taxes, and depreciation and maintenance costs of well pump facilities; reclaimed water includes treatment costs and costs of pressurizing and maintaining the reclaimed water pipeline network; and surface water includes water rights fees and water source engineering maintenance costs. The second objective function aims to minimize the total water shortage, expressed as: To minimize the total water shortage, This represents the total water shortage; among which, Represents grid cells The actual irrigation water requirement, in cubic meters; This represents the total amount of water actually supplied to that grid, i.e. ; The spatiotemporal optimization model strictly follows the key constraints generated by the multimodal large model as hard boundary constraints, and adjusts the priority allocation rule set by multiplying it by different weights according to the actual situation as soft guiding constraints. Among them, the hard boundary constraints include independent water supply capacity constraints, water conveyance capacity constraints, power grid load constraints, and crop-water source adaptability constraints based on water quality matching. The soft guiding constraints are transformed from the priority allocation rule set generated by the multimodal large model, including prioritizing the use of surface water, prioritizing the use of low-cost surface water, limiting the use of groundwater in over-extraction areas, and prioritizing the scheduling of gravity-flowable water sources during peak electricity consumption periods. Finally, based on the objectives of maximizing the overall economic benefits of agricultural irrigation areas and minimizing the total water shortage, the spatiotemporal optimization model integrates hard boundary constraints and soft guiding constraints, and uses a multi-objective evolutionary algorithm to obtain the Pareto optimal solution set for the water allocation method of agricultural irrigation areas.

[0036] Specifically, construct a three-dimensional decision variable matrix. A spatiotemporal optimization model with dual optimization objectives is proposed. This matrix defines the water allocation scheme from all water sources *s* to all grid cells *m* over all time periods *t*. The spatiotemporal optimization model adopts a dual optimization objective framework, simultaneously pursuing the maximization of overall economic benefits and the minimization of total water shortage in the agricultural irrigation area. The first objective function is... Aimed at maximizing economic benefits, its composition comprehensively reflects the benefit-cost structure of irrigation management: the benefit portion is achieved through... Calculations, including crop yield This is the total water supply. The function; the cost component includes the direct cost of water resources. and water transport energy consumption costs Cs accurately represents the inherent cost per unit volume of water from different water sources s. Groundwater includes resource taxes and extraction energy consumption, while reclaimed water includes treatment costs. The formula for water transmission energy cost is... Based on the physical principle of water pumps lifting water against gravity, the process involves substituting the total head along the path, equipment efficiency coefficient, and real-time electricity price into the calculation. This quantifies the energy consumption of the water conveyance process into economic costs, ensuring the physical accuracy and economic integrity of the spatiotemporal model cost accounting. The total head... Different calculations are needed based on the type of water source. For surface water, it is the pumping head; for groundwater, it is the sum of the pumping head and the dynamic water level depth; for reclaimed water, it is the sum of the pumping head and the treatment plant effluent pressure. The second objective function... It aims to minimize the total relative water shortage by calculating the sum of the relative proportions of the supply and demand gaps in each grid unit to characterize the water supply reliability of the entire irrigation district; it transforms the goal of irrigation adequacy into an optimizable mathematical indicator to guide the model to prioritize crop water needs and reduce production risks. The spatiotemporal optimization model strictly adheres to the key constraints generated by the multimodal large model as hard boundary constraints. These include constraints on the independent water supply capacity of each water source to ensure that the water transfer volume does not exceed its own sustainable supply capacity; constraints on water supply capacity to ensure that the water transfer volume does not exceed the sustainable supply capacity of the water source; constraints on water transmission capacity to ensure that the flow rate is within the design capacity of the engineering facilities; constraints on power grid load to prevent power system overload caused by the operation of pumping stations in clusters; and constraints on crop-water source compatibility based on water quality matching, such as prohibiting the allocation of high-salinity reclaimed water to salt-sensitive crop grids. These constraints together ensure the physical feasibility and system safety of the optimization scheme. The spatiotemporal optimization model incorporates the priority allocation rule set output by the multimodal large model as soft guiding constraints into the optimization process. For example, it assigns a lower cost weight to surface water by adjusting the cost coefficient to reflect the rule of prioritizing the use of surface water, and introduces a higher energy price during peak electricity consumption periods to achieve a strategy of avoiding pumping water during peak electricity consumption periods. These soft constraints are flexibly adjusted through weight allocation, transforming expert knowledge and scheduling strategies into mathematical preference guidance. The effect is that the optimization results are not only mathematically optimal, but also more in line with the experience and rules in actual management. Finally, a multi-objective evolutionary algorithm spatiotemporal optimization model is used to solve the problem. This algorithm effectively explores the high-dimensional solution space through the population evolution mechanism and can handle the nonlinear and non-convex characteristics of the objective function and constraints. Its output is a Pareto optimal solution set of the water allocation method for agricultural irrigation areas. Each solution in this set represents an optimal trade-off between economic benefits and water shortage risk, providing decision-makers with a variety of scientific allocation strategies covering different preferences.

[0037] In one embodiment of the present invention, step S6 includes the following steps: After the multimodal large model completes pre-training and fine-tuning, the structured Pareto optimal solution set and real-time multimodal data corresponding to the decision time, as well as the structured analysis report provided by experts, which compares these schemes from multiple dimensions, provides visualization charts and suggestions, and gives contextualized recommendations, are input into the multimodal large model for training. Among them, the key performance indicators of the Pareto optimal solution set need to clearly include the water consumption of each water source, the cost ratio of each water source, the intensity of groundwater extraction, and the utilization rate of reclaimed water for each scheme; the analysis report provided by experts needs to focus on comparing and demonstrating the water source structure, sustainability, and risks of different schemes. After the multimodal large model is trained, the Pareto optimal solution set obtained by solving the spatiotemporal optimization model and the multimodal data collected in real time in the digital twin substrate are input into the multimodal large model and subjected to deep fusion analysis. The multimodal large model generates natural language decision suggestions that include multi-dimensional comparative analysis and visual evidence of water source utilization efficiency, economic efficiency and environmental sustainability, which assists in the selection of the final method for water allocation in agricultural irrigation areas.

[0038] Specifically, after the multimodal large model completes its pre-training and instruction fine-tuning, to enable it to ultimately generate high-quality decision-making suggestions, it needs to undergo final decision-reasoning training. The specific implementation method is as follows: The training sample consists of a structured Pareto optimal solution set and real-time multimodal data corresponding to the decision-making moment, along with a structured analysis report provided by domain experts for this specific scenario. This report provides multi-dimensional comparative analysis of these solutions, includes visual charts and suggestions, and offers contextualized recommendations. The expert analysis report should focus on evaluating the rationality of the water source structure, the sustainability of groundwater extraction, the efficiency of reclaimed water utilization, and the synergistic effect of multiple water sources for different solutions. The training sample is then input into the multimodal large model for training. After the spatiotemporal optimization model obtains the Pareto optimal solution set through a multi-objective evolutionary algorithm, each solution represents a feasible water resource allocation scheme. These schemes, along with the real-time updated multimodal data collected from the digital twin substrate, are input into the already trained multimodal large model. The multimodal large model first performs in-depth analysis and multi-dimensional comparative analysis of the Pareto solution set, calculating the key performance indicators of each scheme, such as total economic benefits, total water shortage, total energy consumption cost, water consumption and proportion of different water sources, groundwater extraction intensity, reclaimed water adaptation rate, crop demand satisfaction rate, and environmental impact indicators. Based on this, the multimodal large model calls its integrated visualization generation module to automatically create rich visual evidence to intuitively display the analysis results, including but not limited to: heat maps of the spatial distribution of water shortage in irrigation areas under different schemes, pie charts of water consumption structure of multiple water sources, spatial distribution maps of groundwater extraction, pie charts of economic benefits, bar charts comparing water consumption of different water sources, radar charts of scheme ranking, and process line graphs of important indicators changing over time. Ultimately, the multimodal large model deeply integrates and analyzes these quantitative analysis results, visual evidence, and water resources knowledge acquired during pre-training and fine-tuning to generate a well-structured and rigorous natural language decision-making recommendation report. This report not only elaborates on the advantages and disadvantages of each alternative, but also comprehensively considers real-time scenarios and long-term strategic goals, proposing preferred recommendations and explaining the reasons. This transforms data-driven optimization results into insights and conclusions that are easy for decision-makers to understand and implement, significantly improving the scientific rigor, transparency, and efficiency of the final solution selection. For example, in drought conditions, the option with the lowest risk of water shortage is selected.

[0039] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1. A method for agricultural irrigation district water allocation based on large model analysis, characterized in that, The method comprises the following steps: S1: divide the agricultural irrigation area into heterogeneous grid units, and fuse the multi-modal static data and dynamic data thereof to construct a digital twin basement; S2: based on the digital twin basement, drive the Penman-Monteith model and the water stress coefficient to calculate the actual irrigation water requirement of each grid, and determine the specific water requirement grid unit; S3: use a shortest path algorithm based on graph theory to analyze the water delivery path from each water source to each water requirement grid unit; use a hydraulic model and a cost quantification model to analyze the energy consumption cost and efficiency loss of each water delivery path; combine the water delivery path and the energy consumption cost and efficiency loss to use an automatic data integration method to construct a supply relationship matrix of the water source-water requirement grid unit; S4: construct a multi-modal large model based on a visual-linguistic architecture, and complete pre-training and fine-tuning; after inputting the real-time data of the digital twin basement and the supply relationship matrix of the water source-water requirement grid unit, the multi-modal large model outputs a set of priority allocation rules and key constraint conditions about the agricultural irrigation water allocation method; S5: construct a spatio-temporal optimization model with double optimization objectives, which strictly follows the key constraint conditions generated by the multi-modal large model as hard boundary constraint conditions, and takes the priority allocation rule set as a soft guide constraint condition; use a multi-objective evolutionary algorithm to solve the spatio-temporal optimization model to obtain a set of Pareto optimal solutions; S6: the multi-modal large model deeply fuses the set of Pareto optimal solutions and the real-time data in the digital twin basement to generate natural language decision suggestions containing multi-dimensional comparative analysis and visualized evidence to assist the selection of the final scheme.

2. The agricultural irrigation district water allocation method based on large model analysis according to claim 1, characterized in that, The step S1 comprises the following steps: The target agricultural irrigation area is divided into a plurality of uniform or non-uniform grid units according to crop types, soil properties, elevations and canal topological structures, and each grid unit is numbered; multi-modal data of each grid unit is collected, including static data and dynamic data; The static data includes crop planting structure, soil water holding capacity, soil property, crop root depth, water pump motor efficiency, upper limit of suitable water content, canal topological structure, special data in the field of water conservancy irrigation and expert demonstration data; The soil property includes wilting coefficient and field water holding capacity; the special data in the field of water conservancy irrigation includes water conservancy engineering drawings, canal topological graph, water level process line graph and corresponding structured description text of drawing report, multi-period remote sensing image and its interpretation report, and supply relationship matrix of water source-water requirement grid unit containing multi-source cost, water quality and engineering property information; the expert demonstration data includes structured analysis report provided by experts for multi-dimensional comparison of agricultural irrigation water allocation scheme, visualized chart suggestion and situational recommendation, and historical agricultural irrigation water scheduling decision cases, each case including multi-modal data input and corresponding expert decision rule output; The dynamic data includes grid unit soil moisture based on remote sensing inversion, crop canopy coefficient, total lift, meteorological data, real-time data near each water source and engineering operation data; The meteorological data includes solar radiation, air temperature, humidity, wind speed, precipitation, and soil heat flux; the real-time data near each water source includes real-time available water quantity, water quality indicators, water extraction cost, pump station lift, and priority settings of different water sources; the engineering operation data includes dynamic water level depth of groundwater, effluent pressure of treatment plant, real-time flow of channel or pipeline, gate opening, pump station on-off state, and energy consumption; The preprocessed multi-modal static data and dynamic data are fused to construct a digital twin basement.

3. The agricultural irrigation district water allocation method based on large model analysis according to claim 1, characterized in that The step S2 includes the following steps: Through the Tetens formula: ； saturated water vapor pressure ; wherein, denotes the temperature; Through the analysis formula: ； Obtaining actual water vapor pressure ; wherein, represents relative humidity, represents saturated water vapor pressure; By Penman-Monteith model standard analysis formula: ; obtaining a reference crop evapotranspiration ; wherein represents solar radiation; represents soil heat flux; represents air temperature; represents wind speed; represents saturated water vapor pressure; represents actual water vapor pressure; The water stress coefficient is obtained by a calculation model When , is 1; when , is 0; when , ; wherein θ represents the grid real-time soil volumetric water content obtained by remote sensing inversion; represents the upper limit of the appropriate water content; represents the wilting coefficient; Through the analysis formula: ； obtaining a theoretical crop water requirement ; wherein, represents a crop coefficient; represents a reference crop evapotranspiration; Through the analysis formula: ； obtaining the actual crop evapotranspiration ; wherein represents the water stress coefficient; represents the theoretical crop water requirement; Through the analysis formula: ； obtaining the soil water storage change amount ; wherein, represents the soil water content in different time periods; represents the main root depth of the crop, in meters; Through solving the soil water balance equation: ； actual irrigation water requirement in cubic meters; wherein, represents actual crop evapotranspiration; represents effective precipitation; represents soil water storage change; represents grid area in square meters; represents grid cell number; By judging whether the actual irrigation water requirement of the grid is greater than 0 If yes, the grid is a water requirement grid unit. If yes, the grid is a water requirement grid unit.

4. The agricultural irrigation district water allocation method based on large model analysis according to claim 3, characterized in that, The step S3 uses a shortest path algorithm based on graph theory to analyze the water delivery path from each water source to each water grid cell, including the following steps: After the specific water grid cell is determined, the specific water grid cell position is determined by the grid cell number; the entire water delivery system is abstracted into a mathematical graph using the network abstraction method; the water source point, gate, pump station, and water grid cell are taken as nodes; the channel section or pipe section connecting the nodes is taken as a variable side, and its weight can be initialized as the physical length of the channel section or pipe section; the shortest water delivery path from any water source node to any water grid cell node is efficiently found by using the shortest path algorithm based on graph theory; and the shortest water delivery path found by calculation is constructed into a graph database.

5. The agricultural irrigation district water allocation method based on large model analysis according to claim 4, characterized in that, The step S3 uses a hydraulic model and a cost quantification model to analyze the energy consumption cost and efficiency loss of each water delivery path; and a supply relationship matrix of water source-water grid cell is constructed using an automated data integration method in combination with the water delivery path and the energy consumption cost and efficiency loss, including the following steps: After the graph database is constructed in the water delivery path analysis, the hydraulic model and the cost quantification model are used to analyze the energy consumption and efficiency loss of each path according to the inherent characteristics of different types of water sources; the core of the energy consumption analysis is a pump station energy consumption calculation model; whether each determined water delivery path contains a pump station facility is traversed based on the graph database, and if there is, the operating parameters of the pump station and the real-time energy price are extracted from the engineering database; then, a water pump energy consumption formula is used to calculate the energy consumption cost of delivering unit water quantity: ； Obtaining energy cost ; wherein, represents water density; represents gravitational acceleration; represents total head; represents pump efficiency; represents motor efficiency; represents electricity price; Total head It is necessary to calculate differentially according to the type of water source. For surface water, it is the pump station head; for groundwater, it is the sum of the pump station head and the dynamic water level depth; for reclaimed water, it is the sum of the pump station head and the effluent pressure of the treatment plant. Secondly, efficiency loss analysis is carried out, the core of which is the canal system water delivery efficiency model; according to the data of the canal section and real-time meteorological data, a water delivery efficiency coefficient between 0 and 1 is given to it ; Through the analysis formula: ； cost of efficiency loss ; wherein represents a water delivery efficiency coefficient; represents a source water price; Water source water price Not a single value, need to determine according to the different water source acquisition and processing cost, surface water price is the basic resource fee, groundwater price needs to include resource tax and ecological compensation cost, reclaimed water price needs to cover the depth processing and transportation cost; Finally, the energy consumption cost and the efficiency loss cost are superimposed to obtain the total cost of delivering unit water quantity from a specific type of water source to a specific water grid cell; After the comprehensive analysis of the water delivery path and its cost, a supply relationship matrix of water source-water grid cell is constructed using an automated data integration method. The matrix is a three-dimensional structure, the row index represents different water sources, the column index represents water grid cells, and the depth dimension includes unit water delivery total cost, path topological information, and water quality applicability index.

6. The agricultural irrigation district water allocation method based on large model analysis according to claim 1, characterized in that, The step S4 includes the following steps: A multi-modal large model based on visual-linguistic architecture is constructed, and the multi-modal large model is pre-trained based on the historical data integrated in the digital twin basement; the training data includes water conservancy engineering drawings, canal system topology maps, water level process line maps, and structured description texts corresponding to drawing reports, multi-period remote sensing images and their interpretation reports, and supply relationship matrices of water source-demand grid cells containing multi-water source cost, water quality and engineering attribute information; after pre-training, the large model deeply learns the professional terms, spatial relationships, physical mechanisms, and characteristics of different water sources and the principle of coordinated utilization in the field of water conservancy irrigation; After the pre-training of the multi-modal large model is completed, the multi-modal large model is further trained by using the instruction fine-tuning and supervised learning method; the training data is derived from the historical scheduling decision cases stored in the digital twin basement, and each case contains multi-modal data input and corresponding expert decision rule output; through fine-tuning training, the multi-modal large model can generate rule texts and constraint conditions in accordance with the expert thinking and applicable to the multi-water source scene according to the input multi-modal data; After the pre-training and fine-tuning training of the multi-modal large model are completed, the multi-modal large model is deployed in the decision loop of the digital twin basement; in actual application, the digital twin basement is used as a data source to input the fused multi-modal dynamic and static data and the supply relationship matrix of the water source-demand grid cell to the multi-modal large model in real time, and the multi-modal large model analyzes the data to output the priority allocation rule set and the key constraint conditions for guiding the water allocation method of the agricultural irrigation area.

7. The agricultural irrigation district water allocation method based on large model analysis according to claim 1, characterized in that, The step S5 comprises the following steps: with a three-dimensional matrix as the core, a space-time optimization model with double optimization targets of maximizing the overall economic benefits and minimizing the total water shortage in agricultural irrigation areas is constructed, wherein, is the water source index, is the grid cell index, is the time period index, that is, the water quantity from the water source to the grid cell in the time period ; The first objective function pursues the maximization of economic benefits: Through analysis formula: ； Cost of energy consumption for water delivery wherein, represents the total water volume delivered, in cubic meters; represents the water density; represents the gravitational acceleration; represents the total head; represents the pump efficiency; represents the motor efficiency; represents the electricity price;​ Total head It needs to be calculated differently according to the type of water source. For surface water, it is the pump station head; for groundwater, it is the sum of the pump station head and the dynamic water level depth; for reclaimed water, it is the sum of the pump station head and the treated water pressure of the treatment plant. By function expression: ; maximize economic benefit, denotes economic benefit; wherein, denotes grid unit crop unit benefit, unit: yuan / kg; denotes yield, unit: kg; denotes unit water cost of water source , unit: yuan / m³; denotes total water quantity diverted from the water source, i.e. ; denotes water delivery energy consumption cost function; Cs is the inherent cost of water source s, wherein, the groundwater includes the exploitation energy consumption cost, resource tax and well pump facility depreciation and maintenance fee; the reclaimed water includes the treatment cost and the reclaimed water pipe network boosting and maintenance cost; the surface water is the water taking right fee and the water source engineering maintenance fee; The second objective function pursues the minimization of total water shortage, and the function expression is: ； to obtain a minimum total water deficit, denotes the total water deficit; wherein, denotes the actual irrigation water requirement of a grid cell in cubic meters; denotes the total water supplied to the grid, i.e. ; The space-time optimization model strictly follows the key constraint conditions generated by the multi-modal large model as hard boundary constraint conditions, and adjusts the priority allocation rule set as soft guiding constraint conditions by multiplying different weights; wherein, the hard boundary constraint conditions include the water source supply capacity constraint, the water transmission capacity constraint, the power grid load constraint and the crop-water source adaptability constraint based on water quality matching of each water source; the soft guiding constraint conditions are converted from the priority allocation rule set generated by the multi-modal large model, including the priority use of surface water, the priority use of low-cost surface water, the limitation of groundwater use in over-exploitation area, and the priority scheduling of gravity self-flowing water source in power peak period; Finally, the space-time optimization model solves the Pareto optimal solution set of the water allocation method of the agricultural irrigation area based on the maximization of the overall economic benefits of the agricultural irrigation area and the minimization of the total water shortage, and comprehensively considers the hard boundary constraint conditions and the soft guiding constraint conditions, and uses a multi-objective evolutionary algorithm to obtain the Pareto optimal solution set of the water allocation method of the agricultural irrigation area.

8. The agricultural irrigation district water allocation method based on large model analysis according to claim 1, characterized in that, The step S6 comprises the following steps: After the multimodal large model completes the pre-training and fine-tuning training, the structured Pareto optimal solution set and real-time multimodal data corresponding to the decision moment, and the structured analysis report provided by experts for multi-dimensional comparison, visualization chart suggestion and situational recommendation of these schemes are input into the multimodal large model for training; wherein the key performance indicators of the Pareto optimal solution set need to clearly include the water yield of each scheme, the cost proportion of the water source, the groundwater exploitation intensity, and the water utilization rate; the analysis report provided by the experts needs to focus on the comparison and demonstration of the water source structure, sustainability and risk of different schemes; After the multimodal large model training is completed, the Pareto optimal solution set obtained by solving the space-time optimization model and the real-time collected multimodal data are input into the multimodal large model and deep fusion analysis is performed; the multimodal large model generates output including natural language decision suggestions containing multi-dimensional comparative analysis and visualization evidence of water source efficiency, economy, environmental sustainability, which assists in the selection of the final method of agricultural irrigation district water allocation.

Images