Efficient prediction method for spatial and temporal distribution of grid-scale underground water level in large-scale agricultural irrigation area

Abstract

The invention provides a large-scale agricultural irrigation area grid scale underground water level spatial and temporal distribution efficient prediction method, and belongs to the technical field of underground water prediction.The method comprises the steps that firstly, an SWAT model of a drainage basin where an irrigation area is located and an MODFLOW model covering the irrigation area are constructed respectively, the SWAT model and the MODFLOW model are coupled to obtain an SWAT-MODFLOW model, and parameter correction is conducted; secondly, an irrigation area underground water level data set is constructed through scene combination and model simulation and used for training and screening the LSTM model with the highest precision, the optimal control range of each LSTM model is determined through space compression and reconstruction, and a multi-LSTM model is obtained; and finally, carrying out parallel calculation on the multiple LSTM models to predict the spatial and temporal distribution of the grid scale underground water level of the irrigation area. According to the method, the optimal control range of each LSTM model is determined through space compression-reconstruction, the prediction of the underground water level spatial and temporal distribution of the grid scale of the large-scale agricultural irrigation area is finally realized through multi-LSTM model parallel calculation, and the prediction efficiency is improved on the premise of ensuring the prediction precision of the underground water level of the irrigation area.

Description

Technical Field

[0001] The present invention belongs to the field of groundwater prediction technology, and relates to an efficient prediction method for the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas, and in particular to an efficient prediction method for the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas based on the coupling of spatial compression, reconstruction and artificial intelligence technologies. Background Art

[0002] China, a country with a large population and a large agricultural sector, relies heavily on irrigation for agricultural production. In some agricultural irrigation areas located in arid and semi-arid regions of northern China, limited surface water resources make it difficult to sustain agricultural production. Groundwater has become the primary source of irrigation water, accounting for over 70% of total water use. Years of large-scale groundwater extraction have led to a continuous decline in groundwater levels within irrigation areas, triggering a series of ecological and environmental problems such as underground funneling, reduced river flow, and ecological degradation. In this context, it is crucial to determine how regional groundwater levels respond to meteorological conditions and various groundwater resource management measures, and then to identify targeted groundwater resource management measures through optimization calculations.

[0003] Currently, there are two main approaches for predicting regional groundwater levels in response to meteorological conditions and groundwater resource management measures. One approach involves physical models, such as MODFLOW (Modular Three-Dimensional Finite-Difference Groundwater Flow Model). These models primarily use finite difference or finite element methods to solve partial differential equations, thereby predicting the spatiotemporal distribution of regional groundwater levels on a grid-by-grid basis. For example, Chinese invention patent (201911057334.1) provides a method for simulating hydrogeological characteristics and regional groundwater circulation for quantitative evaluation of groundwater flow. In this patent, the MODFLOW model is used to simulate regional groundwater dynamics. While the MODFLOW model can accurately calculate regional groundwater levels at a grid-scale, it carries a high computational cost, a disadvantage exacerbated when coupled with optimization algorithms to optimize groundwater resource management measures.

[0004] Another type of machine learning model, represented by LSTM (Long Short-Term Memory) and RNN (Recurrent Neural Network), primarily uses nonlinear mapping to determine the response of regional groundwater levels to meteorological conditions and groundwater resource management measures. The mainstream modeling approach in this category is to train a machine learning model for all sample points (observation wells or grids) within a region and use the trained model to predict groundwater levels at all sample points within the entire region. For example, Chinese invention patent (202510300451.5) provides a training method and related device for a groundwater level prediction model. This patent trains a machine learning model using data from monitoring points within a region and uses the trained model to predict groundwater levels at all monitoring points within the region. Compared to the MODFLOW model, machine learning models offer the advantage of higher computational efficiency. However, due to the heterogeneity of factors such as hydrological conditions, geological conditions, and human activity intensity within large-scale agricultural irrigation areas, groundwater level fluctuation patterns are often inconsistent across regions. This modeling approach, relying solely on a single model, cannot guarantee prediction accuracy at every grid point, and therefore does not achieve satisfactory performance in large-scale agricultural irrigation areas.

[0005] Existing technical methods cannot achieve a balance between groundwater level prediction accuracy and efficiency. How to improve the prediction efficiency of the model while ensuring the accuracy of groundwater level prediction in large-scale agricultural irrigation areas is a problem that needs to be solved. Summary of the Invention

[0006] To address the problems existing in the prior art, the present invention provides a method for efficiently predicting the spatiotemporal distribution of groundwater levels at the grid scale in large-scale agricultural irrigation areas. Starting from the influencing factors and response process of the groundwater level in each grid, the Nash–Sutcliffe Efficiency (NSE), an evaluation method for LSTM prediction accuracy, is used as the main factor to cluster grids with similar physical and hydrological characteristics. By parallel computing of LSTM models established for control grids at the centroid positions of each cluster, efficient and accurate prediction of monthly groundwater levels for all grids in the irrigation area is achieved.

[0007] In order to achieve the above object, the technical solution adopted by the present invention is:

[0008] A method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas is provided. The method is based on coupling spatial compression, reconstruction, and artificial intelligence technologies and includes the following steps:

[0009] Step 1: Build a SWAT (Soil and Water Assessment Tool) model for the watershed where the irrigation district is located. Specifically:

[0010] Step 1.1: Combine relevant data including water system data, DEM (Digital Elevation Model) data, land use data, soil data, and irrigation water use data of the irrigation area, and use the ArcSWAT toolkit to construct a SWAT model of the irrigation area's watershed using ArcGIS software as the platform.

[0011] In step 1.2, based on the measured historical runoff data from the hydrological stations in the irrigation area, the SWAT model parameters in step 1.1 were calibrated using the SUFI-2 (Sequential Uncertainty Fitting, version 2) global search algorithm using the SWAT-CUP software platform, so that the SWAT model output can fit the surface runoff process.

[0012] Step 1.3: Organize the riverbed seepage, precipitation infiltration, and groundwater evaporation data output by the SWAT model in step 1.2 for coupling the SWAT and MODFLOW models in the subsequent step 3.

[0013] Step 2: Build a MODFLOW model covering the irrigation area. Specifically:

[0014] Combining relevant data including borehole data, hydrogeological maps, and hydrogeological parameters, a MODFLOW model covering the irrigation area was constructed with a 1km×1km grid scale using VisualMODFLOW software to output groundwater level simulation results at the irrigation area grid scale.

[0015] Step 3: Couple the SWAT model with the MODFLOW model to obtain the SWAT-MODFLOW model, and perform parameter correction so that the surface runoff and groundwater level output by the SWAT-MODFLOW model can fit the historical measured data. Specifically:

[0016] In step 3.1, the riverbed seepage, precipitation infiltration, and groundwater evaporation output by the SWAT model in step 1.3 are used as input data and input into the MODFLOW model constructed in step 2.1 to drive the simulation of the groundwater level in the irrigation area, completing the coupling of the SWAT model and the MODFLOW model to obtain the SWAT-MODFLOW model.

[0017] Step 3.2: Using the historical measured groundwater level of the observation well as a reference, adjust the hydrogeological parameters of each grid in the SWAT-MODFLOW model constructed in step 3.1 so that the irrigation area groundwater level simulated by the SWAT-MODFLOW model fits the historical measured groundwater level process;

[0018] Step 3.3: If the simulation accuracy index R of the groundwater level in step 3.2 is2 If the coefficient of determination is less than 0.99, the range of values ​​of the parameters related to riverbed seepage, precipitation infiltration, and groundwater evaporation in the SWAT model constructed in step 1 is re-estimated, and the output of the SWAT model is transferred to the MODFLOW model constructed in step 2 again. This cycle is repeated until the surface runoff and groundwater level output by the SWAT-MODFLOW model in step 3.1 can fit the historical measured data. The surface runoff simulation accuracy requires R 2 ≥0.7, groundwater level simulation accuracy requirement R 2 ≥0.99.

[0019] Step 4: Construct multiple meteorological and groundwater extraction scenarios, and construct a groundwater level dataset for the irrigation area through scenario combination and model simulation, which is used to train and screen the LSTM model. Specifically:

[0020] Step 4.1: Combine the historical and future meteorological data of the basin to construct a multi-category meteorological scenario dataset as the input of the SWAT model in the SWAT-MODFLOW model in step 3;

[0021] Step 4.2: Combine historical groundwater extraction data of the irrigation area to construct a dataset of multiple groundwater extraction scenarios, which will be used as input to the MODFLOW model in the SWAT-MODFLOW model in step 3.

[0022] Step 4.3: Process the multi-class meteorological dataset constructed in step 4.1 and the multi-class groundwater extraction dataset constructed in step 4.2 in pairs to construct a combination scenario that intersects the multi-class meteorological scenario and the groundwater extraction scenario. The pairwise combination processing method specifically means that each meteorological scenario is combined with each groundwater extraction scenario to form a groundwater extraction scenario.

[0023] In step 4.4, the various combination scenarios obtained in step 4.3 are used as inputs to the SWAT-MODFLOW model in step 3. The groundwater level response process of each grid in the irrigation area under various combination scenarios is determined by SWAT-MODFLOW model simulation.

[0024] In step 4.5, based on the time span, the groundwater level response process for each grid cell within the irrigation district under each combination of scenarios determined in step 4.4 is divided into a training set, a validation set, and a test set, with a ratio of 70%, 15%, and 15%. The training set is used for parameter training of the LSTM model, the validation set is used for hyperparameter tuning of the LSTM model, and the test set is used to evaluate the prediction accuracy of the LSTM model. Ultimately, the LSTM model that meets the required accuracy (Nash efficiency coefficient NSE ≥ 0.5) and has the highest accuracy on the test set is selected.

[0025] Step 5: Determine the optimal control range of each LSTM model through space compression and reconstruction. Specifically:

[0026] Step 5.1: Based on the spatial distribution of six characteristics (infrastructure coefficient, water supply degree, land use type, elevation, aquifer thickness, and standard deviation of groundwater level fluctuation) for each grid within the irrigation district, the entire irrigation district was clustered into multiple sub-districts using the K-means clustering algorithm and the silhouette coefficient optimization method. The grid at the center of gravity of each sub-district was used as the control grid for that sub-district.

[0027] Step 5.2: Build and train an LSTM model for each partitioned control grid to predict the groundwater level at the control grid under various combination scenarios in step 4.3.

[0028] Step 5.3: Extend the parameters of the LSTM model corresponding to each control grid to all grids in the irrigation area by traversing, determine the accuracy distribution of each LSTM model parameter in predicting the groundwater level of all grids, and preliminarily construct a multi-LSTM model accuracy evaluation matrix. ;

[0029] The multi-LSTM model accuracy evaluation matrix is ​​shown in formula (1):

[0030] (1)

[0031] Among them, the multi-model accuracy evaluation matrix Line label for each line Represents a grid, with the column label of each column Represents a control grid. Row, No. Column values Representative The grid is using The mean accuracy obtained on the test set of various scenarios when the parameters of the control grid model are changed.

[0032] Step 5.4: Extract the multi-LSTM model accuracy evaluation matrix obtained in step 5.3 The maximum value of each row determines the optimal accuracy that can be achieved by using multiple LSTM models to jointly predict the groundwater level of all grids in the irrigation area, and the spatial distribution of the optimal accuracy is determined based on the spatial location of each grid;

[0033] In step 5.5, add a control grid in the area with weak accuracy and repeat steps 5.2 to 5.4 to complete the correction of the multi-LSTM model accuracy evaluation matrix.

[0034] Step 5.6: Determine the optimal control range of each control grid through the precision clustering algorithm. Specifically:

[0035] If the matrix in step 5.5 No. Line in its The column takes the maximum value, then The grid is summarized in The optimal control range of the LSTM model.

[0036] Step 6: Multiple LSTM models are used in parallel to predict the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area. Specifically:

[0037] In the process of predicting the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area, the LSTM model constructed and trained for each control grid is calculated in parallel to obtain the groundwater levels of all grids within the optimal control range of each LSTM model. Combined with the spatial position of each grid, the spatiotemporal distribution of groundwater levels at the grid scale of the entire irrigation area is obtained.

[0038] The beneficial effects of the present invention compared to the prior art are:

[0039] (1) Traditional physical models such as MODFLOW and FEFLOW can realize groundwater level prediction at the grid scale of large-scale agricultural irrigation areas. The calculation results are accurate and have high spatial resolution. However, their main disadvantage is the high time cost required for calculation, which makes it difficult to couple the simulation of groundwater level with the optimization algorithm to optimize the calculation of groundwater resource management measures. The present invention first uses the SWAT-MODFLOW model constructed in steps 1-3 to determine the response process of groundwater level in each grid of the irrigation area under various meteorological conditions and groundwater exploitation conditions, and then uses the groundwater level response process (output) of the irrigation area under various meteorological conditions and groundwater exploitation conditions (input) to train the LSTM model. Finally, the groundwater level process of each grid in the irrigation area is predicted by parallel computing technology (step 6). While ensuring accuracy and resolution, the calculation efficiency is greatly improved.

[0040] (2) Most current artificial intelligence technologies use a single model to cover the entire study area. However, due to the heterogeneity of groundwater level fluctuation patterns caused by differences in physical characteristics and hydrological properties between grids, the model training accuracy is poor. In the actual prediction process, the accuracy of each grid cannot be taken into account, which ultimately leads to poor overall model accuracy or generalization ability. The present invention first uses spatial compression (steps 5.1-5.2) to determine the control grid, and then performs one-to-one modeling for each control grid, thereby avoiding the interference of heterogeneous groundwater level fluctuation patterns on parameter training and significantly improving training accuracy. Spatial reconstruction (steps 5.3-5.6) is then used to find the spatial range with the most similar physical characteristics and hydrological characteristics to the control grid, thereby significantly improving the prediction accuracy of groundwater levels for all grids in the irrigation area to be studied. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of the flow of the method for predicting the spatiotemporal distribution of groundwater levels at the grid scale in large-scale agricultural irrigation areas according to the present invention;

[0042] Figure 2 The structural diagram of the LSTM model constructed for the present invention;

[0043] Figure 3 This is a schematic diagram of the irrigation area grid division according to an embodiment of the present invention;

[0044] Figure 4 This figure shows the simulation results of combined scenarios and the division of groundwater level datasets in an irrigation district according to an embodiment of the present invention. Each broken line in the figure represents the response process of the groundwater level of a single grid under a certain type of combined scenario. The time span is from January 2001 to December 2064. According to the ratio of 70%, 15%, and 15%, the data from January 2001 to October 2045 is divided into the training set, the data from October 2045 to May 2055 is divided into the validation set, and the data from May 2055 to December 2064 is divided into the test set.

[0045] Figure 5 This is a location distribution diagram of the irrigation area control grids according to an embodiment of the present invention. Each black circle in the diagram represents a control grid.

[0046] Figure 6 This is a distribution diagram of the accuracy of the LSTM model for each control grid in the irrigation area in predicting the groundwater level of all grids in the irrigation area according to an embodiment of the present invention. The circles in the figure represent the control grids, the green circle represents the current control grid, and the colors of the remaining grids correspond to the NSE prediction accuracy; Figure 6 (a) shows the accuracy distribution of the LSTM model of control grid 1 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (b) shows the accuracy distribution of the LSTM model of control grid 2 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (c) is the accuracy distribution map of the LSTM model of control grid 3 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (d) is the accuracy distribution map of the LSTM model of control grid 4 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (e) is the accuracy distribution map of the LSTM model of control grid 5 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (f) is the accuracy distribution map of the LSTM model of control grid 6 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (g) is the accuracy distribution map of the LSTM model of control grid 7 in predicting the groundwater level of all grids in the irrigation area; Figure 6(h) is the accuracy distribution diagram of the LSTM model of the control grid 8 in predicting the groundwater level of all grids in the irrigation area; Figure 6 (i) is the accuracy distribution diagram of the LSTM model of the control grid 9 in predicting the groundwater level of all grids in the irrigation area;

[0047] Figure 7 is the optimal control range distribution diagram of each control grid in the irrigation area of the embodiment of the present application, the circles in the figure represent control grids, the green circles represent the current control grid, and the blue grids represent the optimal control range of the current control grid; Figure 7 (a) is the optimal control range distribution diagram of the control grid 1; Figure 7 (b) is the optimal control range distribution diagram of the control grid 2; Figure 7 (c) is the optimal control range distribution diagram of the control grid 3; Figure 7 (d) is the optimal control range distribution diagram of the control grid 4; Figure 7 (e) is the optimal control range distribution diagram of the control grid 5; Figure 7 (f) is the optimal control range distribution diagram of the control grid 6; Figure 7 (g) is the optimal control range distribution diagram of the control grid 7; Figure 7 (h) is the optimal control range distribution diagram of the control grid 8; Figure 7 (i) is the optimal control range distribution diagram of the control grid 9;

[0048] Figure 8 is the accuracy distribution diagram of the prediction result of the groundwater level in the irrigation area of the embodiment of the present application; Figure 8 (a) is the accuracy distribution diagram of the NSE index; Figure 8 (b) is the accuracy distribution diagram of the Mean Absolute Error (MAE) index; Figure 8 (c) is the accuracy distribution diagram of the Root Mean Square Error (RMSE) index;

[0049] Figure 9 is a comparison diagram of the cumulative frequency distribution results of the prediction accuracy of the irrigation area in the embodiment and the prediction accuracy of the control example, wherein the blue line represents the results of the embodiment, and the orange line represents the results of the control example; Figure 9 (a) is a comparison diagram of the cumulative frequency distribution results of the NSE index; Figure 9 (b) is a comparison diagram of the cumulative frequency distribution results of the MAE index; Figure 9 (c) is a comparison diagram of the cumulative frequency distribution results of the RMSE index. DETAILED DESCRIPTION

[0050] As mentioned above, the present invention provides an efficient prediction method for the spatiotemporal distribution of groundwater levels in large-scale agricultural irrigation areas based on spatial compression, reconstruction and artificial intelligence technology. The specific process is as follows: Figure 1 As shown, the method includes the following steps:

[0051] Step 1: Build a SWAT model for the watershed where the irrigation district is located. Specifically:

[0052] Step 1.1: Combine relevant data including water system data, DEM data, land use data, soil data, and irrigation water use data of the irrigation area, and use the ArcSWAT toolkit in ArcGIS software as the platform to construct a SWAT model for the irrigation area's basin to describe the surface hydrological process. The expression is as follows:

[0053] (2)

[0054] Where, represents the final soil moisture content, mm; represents the initial soil moisture content, mm; represents the time step, Representative Daily precipitation, mm; Representative Sky and earth surface runoff, mm; Representative Daily evapotranspiration, mm; Representative Soil infiltration and flow rate per day, mm; Representative Daily underground runoff, mm.

[0055] In step 1.2, based on the measured historical runoff data from the hydrological stations in the irrigation area, the SWAT-CUP software was used as the platform, and the SUFI-2 global search algorithm was used to calibrate the parameters of the SWAT model in step 1.1 so that the model output could fit the surface runoff process.

[0056] Step 1.3: Organize the riverbed seepage, precipitation infiltration, and groundwater evaporation data output by the SWAT model in step 1.2 for coupling the SWAT and MODFLOW models in the subsequent step 3.

[0057] Step 2: Build a MODFLOW model covering the irrigation area. Specifically:

[0058] Combining relevant data including borehole data, hydrogeological maps, and hydrogeological parameters, the VisualMODFLOW software was used as a platform to construct a MODFLOW model covering the irrigation area at a grid scale of 1 km×1 km, which was used to output the groundwater level simulation results at the grid scale of the irrigation area. Figure 3 As shown in Figure 2, the main source of groundwater in the irrigation area to be studied is the Quaternary phreatic aquifer, so the groundwater flow mathematical model is generalized into a heterogeneous, isotropic two-dimensional groundwater unsteady flow in MODFLOW, which is expressed as follows:

[0059] (3)

[0060] Where, To simulate the seepage area, m 2 ; (x, y) is the spatial coordinate of the grid, m; is the free surface of groundwater; is a type of boundary; It is a second-class boundary; For time, is the flow rate per unit area on the flow boundary of the seepage zone, m 3 / d / m; is the initial groundwater level, m; is the river water level, m; is the external normal direction of the boundary; is the permeability coefficient, m / d; is the water storage capacity of the aquifer, l / m; is the gravity feed rate of the unconfined aquifer at the phreatic surface; The water added to or discharged from the water table per unit area per unit time, including precipitation infiltration and evaporation; is the source and sink term of the aquifer, l / d.

[0061] Step 3: SWAT-MODFLOW model coupling and parameter modification, specifically:

[0062] Step 3.1: Transfer the riverbed seepage, precipitation infiltration, and groundwater evaporation output by the SWAT model in step 1.3 to the MODFLOW model constructed in step 2 to drive the simulation of the irrigation area's groundwater level and complete the SWAT-MODFLOW model coupling.

[0063] Step 3.2: Using the historical measured groundwater levels in the observation wells as a reference, adjust the hydrogeological parameters of each grid in the MODFLOW model constructed in step 3.1, so that the groundwater levels in the irrigation area simulated by the model fit the historical measured groundwater levels.

[0064] Step 3.3: If the simulation accuracy index R of the groundwater level in step 3.2 is2 If the coefficient of determination is less than 0.99, the range of values ​​of the parameters related to riverbed seepage, precipitation infiltration, and groundwater evaporation in the SWAT model constructed in step 1 is re-estimated, and the output of the SWAT model is transferred to the MODFLOW model constructed in step 2 again. This cycle is repeated until the surface runoff and groundwater level output by the SWAT-MODFLOW model in step 3.1 can fit the historical measured data (where the surface runoff simulation accuracy requires R 2 ≥0.7, groundwater level simulation accuracy requirement R 2 ≥0.99).

[0065] Step 4: Construct multiple meteorological and groundwater extraction scenarios, and construct a groundwater level dataset for the irrigation area through scenario combination and model simulation, which is used to train and select the LSTM model with the highest accuracy. Specifically:

[0066] In step 4.1, four types of meteorological scenario datasets are constructed by combining the historical and future meteorological data of the basin. These datasets are used as input for the SWAT model in the SWAT-MODFLOW model in step 3. Specifically, the future meteorological scenarios include historical replay and three types of shared socioeconomic pathways (SSPs) under the BCC-CSM2-MR (Beijing Climate Center Climate System Model version 2 Medium Resolution) atmospheric circulation model in the CMIP6 (Coupled Model Intercomparison Project Phase 6) project, namely SSP126, SSP245, and SSP370. The relevant data of each shared socioeconomic pathway scenario are mapped from grid points to four meteorological stations in the basin where the irrigation area to be studied is located after inverse distance weighted interpolation and bias correction.

[0067] In step 4.2, four groundwater extraction scenario datasets were constructed based on the historical groundwater extraction data of the irrigation district. These datasets were used as input for the MODFLOW model in the SWAT-MODFLOW model in step 3. Specifically, the groundwater extraction scenarios were based on the monthly historical groundwater extraction intensity of various land uses in the irrigation district. Four groundwater extraction scenarios were set for each land use type, representing 25%, 50%, 75%, and 100% of the mean historical groundwater extraction intensity.

[0068] Step 4.3: Process the four meteorological datasets constructed in step 4.1 and the four groundwater extraction datasets constructed in step 4.2 in pairs to construct sixteen meteorological scenarios intersecting with groundwater extraction scenarios (the processing method specifically means that each meteorological scenario is combined with each groundwater extraction scenario to form a groundwater extraction scenario);

[0069] In step 4.4, the various combination scenarios obtained in step 4.3 are used as inputs to the SWAT-MODFLOW model in step 3. The groundwater level response process of each grid in the irrigation area under various combination scenarios is determined through model simulation ( Figure 4 );

[0070] Step 4.5: Based on the time span, the groundwater level response process of each grid in the irrigation area under various combination scenarios determined in step 4.4 is divided into training set, validation set and test set in the ratio of 70%, 15% and 15% ( Figure 4 The training set is used for parameter training of the LSTM model, the validation set is used for hyperparameter tuning of the LSTM model, and the test set is used to evaluate the prediction accuracy of the LSTM model and screen the LSTM model with the highest accuracy and meeting the accuracy requirements (NSE ≥ 0.5).

[0071] Step 5: Space compression and reconstruction are performed to determine the optimal control range of each LSTM model. Specifically:

[0072] Step 5.1: Based on the spatial distribution of six characteristics (infrared permeability, water supply, land use type, elevation, aquifer thickness, and standard deviation of groundwater level fluctuation) of each grid in the irrigation area, the entire irrigation area is clustered into multiple sub-areas using the K-means clustering algorithm and the silhouette coefficient optimization method. The grid at the center of gravity of each sub-area is used as the control grid of the sub-area. The location of the control grid is shown in Figure 5. Figure 5 ;

[0073] In step 5.2, a separate LSTM model is established and trained for each control grid in each partition to predict the groundwater level at the control grid under various combination scenarios in step 4.3. The LSTM model takes monthly cumulative precipitation, monthly cumulative evaporation, monthly cumulative groundwater extraction depth, river flow, and groundwater level at the beginning of the month as input, and uses the groundwater level at the end of the month as output. The Adam optimization algorithm is used for parameter training, and the hyperparameters are manually adjusted based on the performance of the parameters in the validation set. The structure of the model is as follows: Figure 2 As shown;

[0074] Step 5.3: Extend the LSTM model parameters corresponding to each control grid to all grids in the irrigation area by traversing, determine the accuracy distribution of each LSTM model parameter in predicting the groundwater level of all grids, and preliminarily construct the multi-LSTM model accuracy evaluation matrix M;

[0075] The multi-LSTM model accuracy evaluation matrix is ​​shown in formula (4):

[0076] (4)

[0077] Among them, the multi-model accuracy evaluation matrix Line label for each line Represents a grid, with the column label of each column Represents a control grid. Row, No. Column values Representative The grid is using The mean accuracy obtained on the test set of various scenarios when the parameters of the control grid model are changed.

[0078] Step 5.4: Extract the multi-LSTM model accuracy evaluation matrix obtained in step 5.3 The maximum value of each row determines the optimal accuracy that can be achieved by using multiple LSTM models to jointly predict the groundwater level of all grids in the irrigation area, and the spatial distribution of the optimal accuracy is determined based on the spatial location of each grid;

[0079] Step 5.5: Add a control grid in the area with weak accuracy and repeat steps 5.2 to 5.4 to complete the correction of the multi-LSTM model accuracy evaluation matrix. The spatial distribution of the optimal accuracy determined by this matrix is ​​as follows: Figure 6 ;

[0080] Step 5.6, determine the optimal control range of each control grid through the precision clustering algorithm. The specific distribution of the optimal control range is as follows: Figure 7 Specifically:

[0081] If the matrix in step 5.5 No. Line in its The column takes the maximum value, then The grid is summarized in The optimal control range of the LSTM model.

[0082] Step 6: Multiple LSTM models are used in parallel to predict the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area. Specifically:

[0083] In the process of predicting the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area, parallel calculations are performed on the LSTM model built and trained for each control grid to predict the groundwater levels of all grids within the optimal control range of each LSTM model. The numerical distribution and cumulative frequency distribution of the groundwater level prediction accuracy are shown in Figure 2. Figure 8 and Figure 9Combined with the spatial position of each grid, the spatiotemporal distribution of groundwater level at the grid scale of the entire irrigation area is obtained.

[0084] This embodiment selects an irrigation area in the Songliao River Basin as a research example, and uses the method described in the present invention to efficiently predict the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area.

[0085] The study irrigation district is located in the Songliao River Basin in Northeast China. The climate in this area is typical of a temperate, semi-arid continental climate, with the rainy season (June to August) accounting for over 80% of annual precipitation. The irrigation district is bounded by the main river channel, and agricultural land predominates within it. The massive expansion of cultivated land in recent years has led to a dramatic increase in water demand. Years of extensive groundwater extraction have led to a continuous decline in groundwater levels within the irrigation district, creating a large-scale water table that encompasses almost the entire area.

[0086] The first step is to construct a SWAT model for the watershed where the irrigation district is located and complete preliminary parameter calibration.

[0087] The second step is to build a MODFLOW model covering the irrigation area with a grid scale of 1 km × 1 km. The grid division results of the irrigation area are as follows: Figure 3 As shown in the figure, there are 1904 grids in total, and the area of ​​each grid is 1 km².

[0088] In the third step, the riverbed seepage, precipitation infiltration, and groundwater evaporation output by the SWAT model are transferred to the MODFLOW model to drive the simulation of the groundwater level, complete the coupling of the SWAT-MODFLOW model, and complete the correction of the SWAT-MODFLOW model parameters based on historical measured runoff data and groundwater level data.

[0089] The fourth step is to determine the response process of the groundwater level of each grid in the irrigation area under various combined scenarios through multi-scenario simulation, and construct a groundwater level dataset at the grid scale of the irrigation area. Here, the effect diagram of the groundwater level dataset and the processing method are drawn using grid No. 125 as an example ( Figure 4 ).

[0090] The fifth step is to gradually determine the position of the control grid through space compression and reconstruction ( Figure 5 ), the accuracy distribution of the LSTM model of each control grid in predicting the groundwater level of all grids in the irrigation area ( Figure 6 ) and the optimal control range of each control grid ( Figure 7 ).

[0091] The sixth step is to achieve efficient and accurate prediction of the spatiotemporal distribution of groundwater levels at the grid scale of the irrigation area through parallel computing of multiple LSTM models. This calculation process is run on a computing server equipped with an Intel(R) Xeon(R) Gold 5218R processor and an NVIDIA A100 GPU. The parallel computing of multiple LSTM models can complete a 116-month prediction in about 13 seconds, which is about 10 times more efficient than the physical model. The prediction accuracy of the model is evaluated using three indicators: NSE, MAE, and RMSE. The numerical distribution of the prediction results with reference to the accuracy of the physical model is shown in the figure. Figure 8 ; Comparison of the accuracy cumulative frequency distribution of the embodiment and the control example is shown in Figure 9 .

[0092] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention and is not limiting. Although the present invention is described in detail with reference to the preferred arrangement scheme, it should be understood by those skilled in the art that the technical solution of the present invention (such as the use of various formulas, the sequence of steps, etc.) can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. An efficient method for predicting the spatiotemporal distribution of groundwater levels at the grid scale in large-scale agricultural irrigation areas, characterized by: The method for efficiently predicting the spatiotemporal distribution of groundwater levels at the grid scale in large-scale agricultural irrigation areas comprises the following steps: Step 1: construct a SWAT model for the watershed where the irrigation district is located; Step 2: construct a MODFLOW model covering the irrigation area; Step 3: Couple the SWAT model with the MODFLOW model to obtain the SWAT-MODFLOW model, and perform parameter correction so that the surface runoff and groundwater level output by the SWAT-MODFLOW model can fit the historical measured data; Step 4: Construct multiple meteorological and groundwater extraction scenarios. Through scenario combination and model simulation, construct an irrigation district groundwater level dataset and divide it into a training set, a validation set, and a test set. The training set is used for parameter training of the LSTM model, the validation set is used for hyperparameter tuning of the LSTM model, and the test set is used to evaluate the prediction accuracy of the LSTM model. Ultimately, the LSTM model with the highest accuracy on the test set is selected. Step 5: Determine the optimal control range of each LSTM model through space compression and reconstruction to obtain a multi-LSTM model; In step 6, multiple LSTM models are used to perform parallel calculations to predict the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area.

2. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1: Construct a SWAT model for the watershed where the irrigation district is located by combining relevant data including water system data, DEM data, land use data, soil data, and water use data of the irrigation district; In step 1.2, based on the measured historical runoff data from the hydrological stations in the irrigation area, the SUFI-2 global search algorithm is used to calibrate the parameters of the SWAT model in step 1.1, so that the SWAT model output can fit the surface runoff process. Step 1.3: Organize the riverbed seepage, precipitation infiltration, and groundwater evaporation data output by the SWAT model in step 1.2 for coupling the SWAT and MODFLOW models in the subsequent step 3.

3. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 2 is characterized in that: The step 2 is specifically as follows: combining data including borehole data, hydrogeological maps, and hydrogeological parameters to construct a MODFLOW model covering the irrigation area, which is used to output groundwater level simulation results at the grid scale of the irrigation area.

4. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 3 is characterized in that: The step 3 is specifically as follows: In step 3.1, the riverbed seepage, precipitation infiltration, and groundwater evaporation output by the SWAT model are used as input data into the MODFLOW model to drive the simulation of the groundwater level in the irrigation area, completing the coupling of the SWAT model and the MODFLOW model to obtain the SWAT-MODFLOW model. Step 3.2: Using the historical measured groundwater level of the observation well as a reference, adjust the hydrogeological parameters of each grid in the SWAT-MODFLOW model so that the groundwater level in the irrigation area simulated by the SWAT-MODFLOW model fits the historical measured groundwater level process; Step 3.3: If the simulation accuracy index R of the groundwater level in step 3.2 is 2 If the value is less than 0.99, the value ranges of the parameters related to riverbed seepage, precipitation infiltration, and groundwater evaporation in the SWAT model constructed in step 1 are re-estimated, and the output results of the SWAT model are again transferred to the MODFLOW model constructed in step 2; this cycle is iterated until the surface runoff and groundwater level output by the SWAT-MODFLOW model can fit the historical measured data.

5. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 4 is characterized in that: In step 3.3, the surface runoff and groundwater level output by the SWAT-MODFLOW model can fit the historical measured data. The specific requirements are: the surface runoff simulation accuracy requirement is R 2 ≥0.7, groundwater level simulation accuracy requirement R 2 ≥0.

99.

6. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 4, characterized in that: The step 4 is specifically as follows: Step 4.1: Combine the historical and future meteorological data of the basin to construct a multi-category meteorological scenario dataset as the input of the SWAT model in the SWAT-MODFLOW model in step 3; Step 4.2: Combine historical groundwater extraction data of the irrigation area to construct a dataset of multiple groundwater extraction scenarios, which will be used as input to the MODFLOW model in the SWAT-MODFLOW model in step 3. Step 4.3: Process the multi-class meteorological dataset constructed in step 4.1 and the multi-class groundwater extraction dataset constructed in step 4.2 in pairs to construct a combined scenario that intersects the multi-class meteorological scenario and the groundwater extraction scenario; In step 4.4, the various combination scenarios obtained in step 4.3 are used as inputs to the SWAT-MODFLOW model in step 3. The groundwater level response process of each grid in the irrigation area under various combination scenarios is determined by SWAT-MODFLOW model simulation. In step 4.5, based on the time span, the groundwater level response process of each grid in the irrigation area under various combination scenarios determined in step 4.4 is divided into a training set, a validation set, and a test set in proportion, and an LSTM model with the highest accuracy on the test set is obtained.

7. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 6, characterized in that: In step 4: In step 4.3, the pairwise combination processing method specifically refers to combining each type of meteorological scenario with each type of groundwater extraction scenario to form a type of groundwater extraction scenario; In step 4.5, the accuracy requirement is met when the Nash efficiency coefficient NSE is greater than or equal to 0.

5.

8. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 6, characterized in that: The specific steps in step 5 are as follows: Step 5.1: Based on the spatial distribution of six characteristics (infrastructure coefficient, water supply degree, land use type, elevation, aquifer thickness, and standard deviation of groundwater level fluctuation) for each grid within the irrigation district, the entire irrigation district was clustered into multiple sub-districts using the K-means clustering algorithm and the silhouette coefficient optimization method. The grid at the center of gravity of each sub-district was used as the control grid for that sub-district. Step 5.2: Build and train an LSTM model for each partitioned control grid to predict the groundwater level at the control grid under various combination scenarios in step 4.

3. Step 5.3: Extend the LSTM model parameters corresponding to each control grid to all grids in the irrigation area by traversing, determine the accuracy distribution of each LSTM model parameter in predicting the groundwater level of all grids, and preliminarily construct the multi-LSTM model accuracy evaluation matrix M; Step 5.4: Extract the multi-LSTM model accuracy evaluation matrix obtained in step 5.3 The maximum value of each row determines the optimal accuracy that can be achieved by using multiple LSTM models to jointly predict the groundwater level of all grids in the irrigation area, and the spatial distribution of the optimal accuracy is determined based on the spatial location of each grid; In step 5.5, add a control grid in the area with weak accuracy and repeat steps 5.2 to 5.4 to complete the correction of the multi-LSTM model accuracy evaluation matrix. Step 5.6: Determine the optimal control range of each control grid using the precision clustering algorithm; specifically: If the matrix in step 5.5 No. Line in its The column takes the maximum value, then The grid is summarized in The optimal control range of the LSTM model.

9. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 8, characterized in that: In step 5.3, the multi-LSTM model accuracy evaluation matrix As shown in formula (1): (1) Among them, the multi-model accuracy evaluation matrix Line label for each line Represents a grid, with the column label of each column Represents a control grid; Row, No. Column values Representative The grid is using The mean accuracy obtained on the test set of various scenarios when the parameters of the control grid model are changed.

10. The method for efficiently predicting the spatiotemporal distribution of groundwater levels at a grid scale in large-scale agricultural irrigation areas according to claim 8, characterized in that: The step 6 is specifically as follows: In the process of predicting the spatiotemporal distribution of groundwater levels at the grid scale in the irrigation area, the LSTM model constructed and trained for each control grid is calculated in parallel to obtain the groundwater levels of all grids within the optimal control range of each LSTM model. Combined with the spatial position of each grid, the spatiotemporal distribution of groundwater levels at the grid scale of the entire irrigation area is obtained.

Images