Complex groundwater environment precise ML quantitative model construction method and device

Abstract

The invention discloses a complex groundwater environment precise ML quantitative model construction method and device, and the method comprises the steps: selecting a correlation characteristic variable related to the spatial-temporal evolution of a complex groundwater pollution environment, and building a complex groundwater environment spatial-temporal evolution conceptual model for a typical condition framework; a proper multi-parameter complex groundwater environment spatio-temporal evolution control equation is established and selected, a corresponding groundwater environment spatio-temporal evolution numerical simulation system is obtained, a typical scene data set is obtained accordingly, and a complex groundwater environment accurate ML quantitative model with different pertinence and application characteristics and certain universality is established according to requirements. According to the invention, reliable quantification and prediction of soil and underground water environment states, evolution processes and related influences in different application scenes can be accurately and efficiently realized on the premise of satisfying a given condition framework; and the accuracy, high efficiency and practicability of prediction, pre-judgment and real-time evaluation of the spatial-temporal evolution state of the complex groundwater pollution environment, the spatial-temporal distribution of pollutants and the like are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of soil-groundwater pollution prediction, environmental assessment and risk control, and specifically to a method and device for constructing an accurate ML quantitative model for a complex groundwater environment. Background Art

[0002] At present, in the prediction and evaluation of groundwater pollution and risk assessment, the application of numerical simulation systems to evaluate and quantify complex groundwater pollution status and environmental risks for specific application scenarios is a common research method. However, for accurate numerical simulation, it is necessary to comprehensively consider and refine the many influencing factors and complex evolution processes of specific scenarios, which is time-consuming and highly professional. In addition, there is a large uncertainty in the modeling and prediction of complex groundwater systems. The use of Monte Carlo and other methods for uncertainty assessment requires hundreds or thousands of numerical simulation tests. However, the computational cost is huge or even unacceptable, and the efficiency is low. Therefore, there is an urgent need for a complex groundwater environment spatiotemporal evolution prediction and quantification method with strong applicability, accuracy, efficiency, simplicity and practicality for different specific conditions and a certain degree of universality.

[0003] In view of the above problems, the present invention provides a method and device for constructing an accurate ML quantitative model for a complex groundwater environment, so as to solve the problems in the prior art of the suitability and accuracy of the models used in the prediction and quantification of characteristic variables of complex groundwater pollution environments under specific conditions, large uncertainty, high cost, low efficiency, and poor real-time performance. Summary of the invention

[0004] The purpose of the present invention is to provide a method for constructing an accurate ML quantitative model of a complex groundwater environment, including selecting associated characteristic variables related to the spatiotemporal evolution of a complex groundwater pollution environment, establishing a conceptual model of the spatiotemporal evolution of a complex groundwater environment for a typical condition framework, establishing and selecting an appropriate multi-parameter complex groundwater environment spatiotemporal evolution control equation, obtaining a corresponding groundwater environment spatiotemporal evolution numerical simulation system, thereby obtaining a typical scenario data set, and establishing an accurate ML quantitative model of a complex groundwater environment with different pertinence and application characteristics and a certain degree of versatility according to demand. The specific steps of the method are as follows:

[0005] Step 1: Determine the set of common characteristics of a type of complex groundwater pollution environment system to be studied, which is called the system condition framework. Corresponding to the system condition framework, give a conceptual model of the spatiotemporal evolution of the complex groundwater pollution environment, select the characteristic variables associated with the spatiotemporal evolution of groundwater pollution, obtain the mathematical control equations of the spatiotemporal evolution of groundwater pollution, and obtain the numerical simulation system of the spatiotemporal evolution of groundwater pollution;

[0006] Step 2: Determine the input parameter set V = [v1, v2, ...v n], select the range of each model input parameter variable to form the model input parameter variable range set U = [u1, u2, ..., u n ];

[0007] Step 3: Based on the range set of model input parameter variables, different typical scenarios consisting of different combinations of model input parameters of the groundwater pollution environment system are obtained through scenario sampling, and corresponding simulation values ​​of characteristic variables associated with different typical scenarios are obtained based on the numerical simulation system. The basic modeling data set is obtained according to different combinations of model input parameters for different typical scenarios and their corresponding numerical simulation values;

[0008] Step 4: When it is necessary to strengthen the modeling data set, the input parameter variables of each model are sampled within their variation range according to the given preset number of supplementary scenarios, and different supplementary scenarios composed of different model input parameter combinations of the groundwater environment system are obtained. Based on the numerical simulation system, the corresponding associated characteristic variable simulation values ​​of the different supplementary scenarios are obtained, and the enhanced modeling data set is obtained by combining the model input parameter combinations of all different supplementary scenarios and their corresponding associated characteristic variable simulation values;

[0009] Step 5: Consider all model input parameters for the associated characteristic variables and complete the comprehensive identification of the main controlling factors;

[0010] Step 6: Partially or fully construct in advance machine learning prediction models with certain generality and different suitability and application characteristics as needed, including ML model 1, ML model 2, ML model 3, ML model 4 and ML model 5;

[0011] Step 7: According to the preset value of ML model accuracy, determine whether the ML modeling model accuracy meets the model target accuracy. If so, complete the modeling and obtain the optimized ML model; otherwise, by reducing the preset model accuracy, increasing scenario sampling, and changing the machine learning algorithm, the final ML modeling model accuracy meets the model target accuracy.

[0012] Step 8. Given an actual application scenario, determine the typical condition framework that meets the given requirements, and determine the appropriate machine learning prediction model mentioned above to further obtain the specific prediction value of the application scenario.

[0013] Furthermore, the selected characteristic variables associated with the spatiotemporal evolution of groundwater pollution in step one include groundwater pollutant concentration, saturated zone pollutant cross-section or interface flux, pollution plume migration and diffusion front position, pollution plume stabilization time, stable pollution plume distribution area, distance between the stable pollution plume front and the pollution source, stable pollution plume average concentration and pollution plume dissipation time, soil pollutant concentration, vadose zone gas phase pollutant concentration, vadose zone cross-section pollutant flux, vadose zone-saturated zone pollutant interface flux, soil-air interface pollutant flux and rock-air interface pollutant flux, and NAPL pollutant spatiotemporal distribution.

[0014] Furthermore, the set of common characteristics of a type of groundwater pollution environmental system to be studied in step one is called the system condition framework, which includes determining the characteristics of atmospheric precipitation infiltration and recharge, determining the characteristics of the vadose zone, determining the characteristics of the saturated zone, determining the characteristics of the pollution source, determining the characteristics of groundwater pollution prevention and control and remediation, determining the characteristics of the source and sink of pumping and injection wells, determining the type of groundwater pollutant migration and transformation process, determining the characteristics of the boundary conditions and determining the characteristics of the initial conditions.

[0015] Furthermore, the comprehensive identification of the main controlling factors in step five includes the application of two or more main controlling factor identification methods, including different machine learning methods, hierarchical analysis, variance analysis, range analysis, and sensitivity analysis, and the "and" or "or" set of the main controlling factors obtained based on the adopted methods to obtain the final main controlling factor parameters.

[0016] Furthermore, the associated characteristic variables and model input parameter variables in step 1 and step 2 are processed by any one or any combination of preset data processing methods, wherein the preset data processing methods include retaining the value as it is, increasing the value by the same multiple, decreasing the value by the same multiple, converting the value logarithmically, unifying the positive and negative values, and normalizing the value, and the model input parameter variables obtain a preset number of scenario samples, including obtaining a preset number of random scenario samples based on the random value of the model input parameter variables according to the variation range of the model input parameter variables, and including selecting the variation level of each of the model input parameter variables and then combining the model input parameter variables to form a deterministic scenario sample;

[0017] Furthermore, the machine learning algorithms in step 7 include KAN and FCN.

[0018] Furthermore, the SVM machine learning optimization modeling method in the application of SVM machine learning modeling principles and optimization modeling methods in step eight includes the optimization of SVM machine learning model modeling optimization methods and the optimization of modeling kernel function types and their associated parameters.

[0019] Furthermore, ML model 1 in step six is ​​an efficient and accurate full-factor model, including all parameters required when the model is applied, and no random occlusion sampling data enhancement during modeling; ML model 2 is a simple and reliable master factor model, including all master factor parameters required when the model is applied, and no random occlusion sampling data enhancement during modeling; ML model 3 is a wide-ranging model in which non-master factor parameters are optional, including all master factor parameters required when the model is applied, but non-master factor parameters can be missing, and the model has good wide adaptability; random occlusion sampling is implemented for data enhancement of non-master factor parameters during modeling; ML model 4 is a wide-ranging model in which any factor parameter can be missing, including any factor parameter can be missing when the model is applied, and the model has good wide adaptability; random occlusion sampling is implemented for data enhancement of all parameters during modeling; ML model 5 is a customized input parameter specific model, including the need to pre-select parameters when the model is applied, and no random occlusion sampling data enhancement during modeling.

[0020] Furthermore, the characteristics of atmospheric precipitation infiltration recharge are determined, including the presence or absence of atmospheric precipitation infiltration recharge, and the spatiotemporal variation characteristics of atmospheric precipitation infiltration recharge intensity; the characteristics of the vadose zone are determined, including the medium type, lithological structure, medium heterogeneity and anisotropy characteristics; the characteristics of the saturated zone are determined, including the medium type, water-bearing system structure, seepage dimension, flow state, medium heterogeneity and anisotropy characteristics; the characteristics of the pollution source are determined, focusing on the phase state of pollutants, the type of pollutants, the spatial distribution characteristics of the pollution source, the source strength variation and dynamic characteristics; the characteristics of groundwater pollution prevention, control and remediation are determined, including the presence or absence of groundwater pollution prevention, control and remediation measures, the type of groundwater pollution prevention, control and remediation measures, and the underground The spatiotemporal distribution characteristics of water pollution prevention, control and remediation measures, determination of the source and sink characteristics of pumping and injection wells include the presence or absence of pumping and injection wells, the spatial distribution characteristics of pumping and injection wells, and the dynamic characteristics of water volume in pumping and injection wells; the determination of the interaction characteristics between surface water and groundwater include the presence or absence of interaction between the two, the spatial distribution characteristics of surface water bodies, and the dynamic characteristics of surface water levels; determination of the types of groundwater pollutant migration and transformation processes include advection processes, diffusion processes, adsorption and desorption processes, degradation processes, volatilization processes, etc.; determination of boundary condition characteristics include the boundary condition characteristics of the hydrodynamic field and the boundary condition characteristics of the hydrochemical field; determination of initial condition characteristics include the initial condition characteristics of the hydrodynamic field and the initial condition characteristics of the hydrochemical field.

[0021] A device for constructing a precise ML quantitative model of a complex groundwater environment, the device includes a conceptual model construction module for the spatiotemporal evolution of a complex groundwater pollution environment with multiple parameters, a comprehensive identification module for main controlling factors, a scenario sampling parameter combination construction and scenario numerical simulation module, a data enhancement module, a feature dimension reduction module, and a precise ML model construction and application module; the data enhancement module includes parameter random occlusion sampling data enhancement; the feature dimension reduction module includes a convolutional neural network; the precise ML model construction and application module includes the construction and application of a precise ML model based on KAN.

[0022] The present invention has the following advantages: the present invention can accurately and efficiently realize the reliable quantification and prediction of the soil and groundwater environmental status, evolution process and related impacts in different application scenarios under the premise of meeting a given condition framework, and significantly improve the accuracy, efficiency and practicality of the prediction and prejudgment and real-time evaluation of the spatiotemporal evolution state of complex groundwater pollution environment and the spatiotemporal distribution of pollutants, so as to provide strong support for the protection of soil and water environment system and the safe utilization of resources, especially the rapid prediction, early warning and reliable risk control of complex groundwater environment system. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 This is a conceptual model diagram of the present invention. DETAILED DESCRIPTION

[0024] In this embodiment, the model adopts a study area size of 20*20*10.5m in length, width and height. It is covered by 1 to 5 layers of heterogeneous soil layers (the number of layers is one of the random variables) and a fractured rock structure is located below. The study area has groundwater at a certain depth as a pollution source of constant concentration. There are periodic air pressure fluctuations and rainfall on the surface. The designed conceptual model is as follows: (a) The lower part of the study area is fractured rock. Three groups of orthogonal fractures are designed and their directions follow the Fisher distribution. The inclination angles of the fractures can be changed. The coupling of fractures and rocks is realized by the UDFM algorithm. The upper part of the study area is covered with layered heterogeneous soil of different thicknesses. (b) The surface is regarded as a surface air layer with a thickness of 0.5m, which has a sinusoidally varying pressure boundary to simulate the periodic variation of air pressure. Different rainfall infiltration values ​​are assigned to the soil surface each quarter. The study area is under a saturated aquifer that flows from west to east and has different hydraulic gradients. (c) The saturated aquifer is a pollution source with a constant VOCs concentration. The initial concentrations in other areas are all 0. It is assumed that the pollutant concentration at the surface air boundary is constant at 0, that is, the surface has the maximum diffusion flux. (e) The model considers the distribution between gas, liquid, and solid phases, but does not consider the attenuation of gas phase VOCs by biochemical reactions, and does not involve the transmission of NAPL phases. (f) The model takes g = 9.8 m / s2 to consider the density effect and the influence of temperature. The time step varies within the entire simulation time range, and the gas phase viscosity is constant. It is assumed that there is a local chemical equilibrium when VOCs are distributed between different phases (gas, water, and solid).

[0025] In this embodiment, a numerical simulation of VOCs pollution in groundwater with a double-layer structure of layered heterogeneous soil and fractured rock is established, and various actual conditions such as surface air pressure fluctuations and rainfall are fully considered to achieve a true characterization of the site; through a detailed investigation of the pollutant migration pattern, a VOCs migration and transformation model under complex medium conditions is established by coupling multiple processes, and a supercomputer is used to establish a numerical model under hundreds of groups of random parameter distributions, and this is used as a basic training set. Through variance analysis, data enhancement based on main control factors, convolution feature extraction, etc., a dimensionality reduction operation of high-dimensional feature data is achieved, and an efficient and reliable deep learning general substitution model is established with stable flux as a label, which solves the problems of dimensional disaster, low versatility and insufficient data volume in the construction of complex groundwater pollution substitution models, and can realize a substitution model scheme with high prediction accuracy when some parameters are input. In particular, a suitable ML model with different characteristics and advantages can be established according to actual needs to provide a reliable ML model for accurate and fast VOCs surface flux prediction.

[0026] In this embodiment, the conceptual model construction and numerical simulation system selection are as follows: Figure 1 As shown in the figure, the model adopts a study area size of 20*20*10.5m in length, width and height. It is covered by 1 to 5 layers of heterogeneous soil layers (the number of layers is one of the random variables) and the bottom is a fractured rock structure. The study area has groundwater at a certain depth, which serves as a pollution source of constant concentration. There are periodic air pressure fluctuations and rainfall on the surface. The designed conceptual model is as follows: (a) The lower part of the study area is fractured rock. Three groups of orthogonal fractures are designed and their directions follow the Fisher distribution. The inclination of the fractures can be changed. The UDFM algorithm is used to realize the coupling of fractures and rocks. The upper part of the study area is covered with layered heterogeneous soil of different thicknesses. (b) The surface thickness of 0.5m is regarded as the surface air layer, with a sinusoidally changing pressure boundary to simulate the periodic change of air pressure, and at the same time according to the seasons. Different rainfall infiltration values ​​are assigned to the soil surface. The study area is under a saturated aquifer that flows from west to east and has different hydraulic gradients. (c) The saturated aquifer is a pollution source with a constant VOCs concentration. The initial concentrations in other areas are all 0. It is assumed that the pollutant concentration at the surface air boundary is constant at 0, that is, the surface has the maximum diffusion flux. (e) The model considers the distribution between gas, liquid, and solid phases, but does not consider the attenuation of gas phase VOCs by biochemical reactions, and does not involve the transmission of NAPL phase. (f) The model takes g = 9.8 m / s2 to consider the density effect and the influence of temperature. The time step varies within the entire simulation time range, and the gas phase viscosity is constant. It is assumed that there is a local chemical equilibrium when VOCs are distributed between different phases (gas, water, and solid).

[0027] In this embodiment, the control equation of the spatiotemporal evolution of VOCs multiphase is characterized by the mass conservation equation (concentration form). FEHM is selected as the numerical simulation system of this embodiment. This numerical simulation system can simulate the flow and transmission of groundwater in fractures and porous media, and is suitable for systems with complex geometries. The finite volume method and finite element method are used to solve the heat and mass conservation equations. FEHM has been widely used and rigorously verified in many fields, including non-isothermal multiphase flow, migration and transformation of gaseous pollutants, and other related fields.

[0028] In this embodiment, the numerical simulation parameters are selected as shown in Table 1. The study takes into account as much as possible the migration and transformation processes that affect VOCs in loose pores and bedrock cracks, and accurately quantifies key indicators such as flux and concentration. A total of 76 parameters are incorporated into the model, including the total soil thickness h, the total number of soil layers l, the thickness of the first to fifth layers of soil h1-h5, the gas-solid partition coefficient K of the first to fifth layers of soil ssa1 -K ssa5 , the soil adsorption coefficient K from the first to the fifth layer sd1 -K sd5 , the soil dispersivity from the first to the fifth layer α s1 -α s5 , the soil permeability K from the first to the fifth layer s1 -K s5 , the soil porosity from the first to the fifth layer p s1 -p s5 , residual saturation of soil from the first to the fifth layer rr1 -s rr5 , the coefficients VG of the soil VG model from the first to the fifth layer a1 -VG a5 , the coefficients VG of the soil VG model from the first to the fifth layer n1 -VG n5 , gas phase diffusion coefficient D v , rock gas-solid partition coefficient K rsa , rock adsorption coefficient K rd , water phase diffusion coefficient D w , Henry coefficient K H , initial concentration of pollutants C0, air pressure fluctuation period A fp , water table depth H, temperature T, rainfall in the first and fourth quarters R 14 , rainfall in the second and third quarters R 23 , hydraulic gradient i, bedrock dispersion α rx (x direction), crack diffusion α fx (x direction), crack density F d , vertical crack angle F vephi , the inclination angle F of the vertical crack 1 vdip1 , the inclination angle F of the vertical crack 2udip2 , horizontal crack inclination F hdip1 , horizontal crack inclination F hephi , ekappa, the mean radius of crack clusters 1 to 3 F r1 -F r3 , the standard deviation F of the radius distribution of crack clusters 1 to 3 rv1 -F rv3 , mean crack width F a , standard deviation of crack width F av . After fully investigating the distribution of each parameter under actual environmental information, each parameter is randomly sampled within a reasonable range, and the model simulation is performed until the surface gas phase pollutant diffusion flux reaches stability. This embodiment is designed to generate 400 sets of models with completely random parameters. In order to run numerical simulations of such a scale, the study uses a supercomputing platform to help realize the calculation and storage of large-scale numerical simulations. The grid generation is performed on the cloud supercomputing platform, and the FEHM solution work is run in the cloud server. The total amount of data generated by the simulation exceeds 20T.

[0029] In this embodiment, basic simulation data acquisition and data enhancement: the 400 groups of numerical simulation data are divided into 350 training sets and 50 validation sets, and data enhancement is performed separately. First, ANOVA is used to extract the main controlling factors of stable flux, stable time, and cumulative amount before stability, and the original data is upsampled by random mask operation to ensure that the main controlling factor of the current label remains unchanged. 5%-50% of the remaining features are randomly deleted, and the 350 training sets are enhanced to 3500 sets, and the 50 validation sets are enhanced to 200 sets. On the basis of determining the main factors, the data enhancement effect of all 76 features participating in random masking and retaining the random mask of the main controlling factors is obtained.

[0030] In this embodiment, data preprocessing: The training of regression tasks using deep neural networks usually requires a data set with sufficiently large known features and labels. The study generated 400 sets of data sets containing 76 features through numerical simulation, with stable flux as the label. In order to eliminate the influence of data dimension, the study used Ln to standardize the input and output. The ln1p standardization function is used for the characteristic parameters to ensure the validity of the data. For the stable flux f, due to its large order of magnitude difference and the negative number after using ln, the study takes the opposite number.

[0031] In this embodiment, the basic architecture of the ML model is constructed by combining KAN with weighted random upsampling. First, a CNN encoder is used for feature extraction, the original fully connected layer is deleted, and the local features are extracted using the convolution layer. The features are extracted layer by layer through multi-layer convolution. The low layer can learn simple local features, while the high layer can combine these local features to form more complex global features. At the same time, the pooling layer is used to reduce the feature dimension, retain important information and reduce the computational complexity, thereby improving the efficiency of the model. Finally, the KAN network is used to complete the regression task, and the number of parameters is greatly reduced. This network structure is referred to as AUCK (ANOVA-UPSAMPLING-CNN-KAN).

[0032] In this embodiment, the main controlling factors are identified as shown in Table 2.

[0033] In this embodiment, the corresponding model is established based on the basic data set: In order to discuss the performance of the network on the original data set, the CNN encoder is used to extract features from the original data, and then enter the KAN network for prediction tasks. The data set contains 76 feature dimensions, the label is stable flux, 350 sets of data are used for training, and 50 sets of data are used for verification. According to the data scale, a two-layer convolutional network is established. The input channel of the first layer is 1, the input sequence length is 76, the number of convolution kernels is 16, that is, the output channel is 16, the convolution kernel size is 3, and the padding and stride are both 1; the second layer uses 32 convolution kernels of size 3, which are directly connected to the KAN network. The verification R of the stable flux is calculated 2 are 0.96 respectively. The specific results are shown in Table 3.

[0034] In this embodiment, two different models are obtained based on data enhancement: ML model 3 and ML model 4. The results are shown in Table 4. The model using weighted random upsampling is better than the model using full parameter random upsampling, which verifies R 2 The performance on stable flux labels has been improved, indicating that under the same network structure, weighted random upsampling based on variance analysis and hierarchical analysis has a good improvement on model accuracy.

[0035] In this embodiment, the specific conditions of the actual application scenario are obtained, and after comparative analysis, it is determined that it meets the given typical condition framework. Since the specific values ​​of all main control factor parameters can be determined, while the values ​​of some non-main control factor parameters are highly uncertain or currently unknown, ML model 2 is selected, and based on the specific values ​​of the actual main control factor parameters, the specific prediction value of the application scenario is further obtained.

[0036] Table 1 Random sampling scheme of model parameters

[0037] Table 2 Results of variance analysis and hierarchical analysis weight discrimination of main control factors

[0038] Table 3. Deep learning network model based on KAN application raw data

[0039] Table 4. Different deep learning network models enhanced by KAN data

[0040] Although the specific embodiments of the present invention are described in detail in conjunction with the accompanying drawings, it should not be understood as limiting the scope of protection of the present invention. Within the scope described in the claims, various modifications and variations that can be made by those skilled in the art without creative work still belong to the scope of protection of the present invention.

[0041] Table 1

[0042]

[0043]

[0044] Table 2

[0045]

[0046] Table 3

[0047]

[0048] Table 4

[0049]

Claims

1. A method for constructing an accurate ML quantitative model of a complex groundwater environment, including selecting associated characteristic variables related to the spatiotemporal evolution of a complex groundwater pollution environment, establishing a conceptual model of the spatiotemporal evolution of a complex groundwater environment based on a typical condition framework, establishing and selecting an appropriate multi-parameter complex groundwater environment spatiotemporal evolution control equation, obtaining a corresponding groundwater environment spatiotemporal evolution numerical simulation system, and obtaining a typical scenario data set based on this, and establishing an accurate ML quantitative model of a complex groundwater environment with different pertinence and application characteristics and a certain degree of versatility according to demand, characterized by: The specific steps of the method are as follows: Step 1: Determine the set of common characteristics of a type of complex groundwater pollution environment system to be studied, which is called the system condition framework. Corresponding to the system condition framework, give a conceptual model of the spatiotemporal evolution of the complex groundwater pollution environment, select the characteristic variables associated with the spatiotemporal evolution of groundwater pollution, obtain the mathematical control equations of the spatiotemporal evolution of groundwater pollution, and obtain the numerical simulation system of the spatiotemporal evolution of groundwater pollution; Step 2: Determine the input parameter set V = [v1, v2, ...v n ] Select the range of each model input parameter to form the model input parameter range set U = [u1, u2, ..., u n ]; Step 3: Based on the range set of model input parameter variables, different typical scenarios consisting of different combinations of model input parameters of the groundwater pollution environment system are obtained through scenario sampling, and corresponding simulation values ​​of characteristic variables associated with different typical scenarios are obtained based on the numerical simulation system. The basic modeling data set is obtained according to different combinations of model input parameters for different typical scenarios and their corresponding numerical simulation values; Step 4: When it is necessary to strengthen the modeling data set, the input parameter variables of each model are sampled within their variation range according to the given preset number of supplementary scenarios, and different supplementary scenarios composed of different model input parameter combinations of the groundwater environment system are obtained. Based on the numerical simulation system, the corresponding associated characteristic variable simulation values ​​of the different supplementary scenarios are obtained, and the enhanced modeling data set is obtained by combining the model input parameter combinations of all different supplementary scenarios and their corresponding associated characteristic variable simulation values; Step 5: Consider all model input parameters for the associated characteristic variables and complete the comprehensive identification of the main controlling factors; Step 6: Partially or fully construct in advance machine learning prediction models with certain generality and different suitability and application characteristics as needed, including ML model 1, ML model 2, ML model 3, ML model 4 and ML model 5; Step 7: According to the preset value of ML model accuracy, determine whether the ML modeling model accuracy meets the model target accuracy. If so, complete the modeling and obtain the optimized ML model; otherwise, by reducing the preset model accuracy, increasing scenario sampling, and changing the machine learning algorithm, the final ML modeling model accuracy meets the model target accuracy. Step 8. Given an actual application scenario, determine the typical condition framework according to specific conditions, and determine the appropriate machine learning prediction model required above to further obtain the specific prediction value of the application scenario.

2. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The selected characteristic variables associated with the spatiotemporal evolution of groundwater pollution in the step one include groundwater pollutant concentration, saturated zone pollutant cross-section or interface flux, pollution plume migration and diffusion front position, pollution plume stabilization time, stable pollution plume distribution area, distance of stable pollution plume front from pollution source, stable pollution plume average concentration and pollution plume dissipation time, soil pollutant concentration, vadose zone gas phase pollutant concentration, vadose zone cross-section pollutant flux, vadose zone-saturated zone pollutant interface flux, soil-air interface pollutant flux and rock-air interface pollutant flux and NAPL pollutant spatiotemporal distribution.

3. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The set of common characteristics of a type of groundwater pollution environment system to be studied in the step one is called the system condition framework, which includes determining the characteristics of atmospheric precipitation infiltration and recharge, determining the characteristics of the vadose zone, determining the characteristics of the saturated zone, determining the characteristics of the pollution source, determining the characteristics of groundwater pollution prevention and control and remediation, determining the source and sink characteristics of pumping and injection wells, determining the source and sink characteristics of pumping and injection wells, determining the type of groundwater pollutant migration and transformation process, determining the boundary condition characteristics and determining the initial condition characteristics.

4. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The comprehensive identification of the main controlling factors in step five includes applying two or more main controlling factor identification methods, including different machine learning methods, hierarchical analysis, variance analysis, range analysis, and sensitivity analysis, and obtaining the final main controlling factor parameter variables based on the "union" or "intersection" set of the main controlling factors obtained by the adopted methods.

5. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The associated characteristic variables and model input parameter variables in step one and step two are processed by any one of the preset data processing methods and any combination thereof, wherein the preset data processing methods include retaining the values ​​as they are, increasing the values ​​by the same multiple, decreasing the values ​​by the same multiple, converting the values ​​into logarithms, unifying the positive and negative values, and normalizing the values; the model input parameter variables are used to obtain a preset number of scenario samples, including obtaining a preset number of random scenario samples based on the range of change of the model input parameter variables and randomly taking values ​​of the model input parameter variables, as well as deterministic scenario samples formed by selecting the change level of each of the model input parameter variables and then combining the model input parameter variables.

6. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The machine learning algorithms in step seven include KAN and FCN.

7. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The SVM machine learning optimization modeling method in the application of SVM machine learning modeling principles and optimization modeling methods in step eight includes the optimization of SVM machine learning model modeling optimization methods and the optimization of modeling kernel function types and their associated parameters.

8. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 1, characterized in that: The ML model 1 in step 6 is an efficient and accurate full-factor model, including all the parameters required when the model is applied, and no random occlusion sampling data is enhanced during modeling; the ML model 2 is a simple and reliable main control factor model, including all the main control factor parameters required when the model is applied, and no random occlusion sampling data is enhanced during modeling; the ML model 3 is a wide-adaptability model in which non-main control factor parameters are optional, including all the main control factor parameters required when the model is applied, but non-main control factor parameters can be missing, and the model has good wide adaptability; random occlusion sampling is implemented for data enhancement of non-main control factor parameters during modeling; the ML model 4 is a wide-adaptability model in which any factor parameter can be missing, including any factor parameter can be missing when the model is applied, and the model has good wide adaptability; During modeling, all parameters are subjected to random occlusion sampling for data enhancement; the ML model 5 is a customized input parameter specific model, including the need to pre-select parameters when the model is applied, and there is no random occlusion sampling data enhancement during modeling.

9. The method for constructing an accurate ML quantitative model for a complex groundwater environment according to claim 3, characterized in that: The determination of atmospheric precipitation infiltration recharge characteristics includes whether there is atmospheric precipitation infiltration recharge and the spatiotemporal variation characteristics of atmospheric precipitation infiltration recharge intensity; the determination of aeration zone characteristics includes medium type, lithological structure, medium heterogeneity and anisotropy characteristics; the determination of saturated zone characteristics includes medium type, water-bearing system structure, seepage dimension, flow state, medium heterogeneity and anisotropy characteristics; the determination of pollution source characteristics focuses on pollutant phase state, pollutant type, spatial distribution characteristics of pollution sources, source strength changes and dynamic characteristics; the determination of groundwater pollution prevention, control and remediation characteristics includes whether there are groundwater pollution prevention, control and remediation measures, types of groundwater pollution prevention, control and remediation measures, groundwater The spatiotemporal distribution characteristics of pollution prevention, control and remediation measures; the determination of the source and sink characteristics of pumping and injection wells includes the presence or absence of pumping and injection wells, the spatial distribution characteristics of pumping and injection wells, and the dynamic characteristics of water volume in pumping and injection wells; the determination of the interaction characteristics between surface water and groundwater includes the presence or absence of interaction between the two, the spatial distribution characteristics of surface water bodies, and the dynamic characteristics of surface water levels; the determination of the types of groundwater pollutant migration and transformation processes includes advection processes, diffusion processes, adsorption and desorption processes, degradation processes, volatilization processes, etc.; the determination of boundary condition characteristics includes the boundary condition characteristics of the hydrodynamic field and the boundary condition characteristics of the hydrochemical field; the determination of initial condition characteristics includes the initial condition characteristics of the hydrodynamic field and the initial condition characteristics of the hydrochemical field.

10. A device for constructing an accurate ML quantitative model for complex groundwater environments, characterized by: The device includes a conceptual model construction module for the spatiotemporal evolution of a complex groundwater pollution environment with multiple parameters, a comprehensive identification module for main control factors, a scenario sampling parameter combination construction and scenario numerical simulation module, a data enhancement module, a feature dimension reduction module, and an accurate ML model construction and application module; the data enhancement module includes parameter random occlusion sampling data enhancement; the feature dimension reduction module includes a convolutional neural network; the accurate ML model construction and application module includes the construction and application of an accurate ML model based on KAN.

Images