Underground water pollution general prediction model construction method and device based on ANN and SVM

Abstract

The invention discloses an underground water pollution general prediction model construction method and device based on ANN and SVM, and the method comprises the steps: selecting a correlation characteristic variable related to the spatial-temporal evolution of underground water pollution, an underground water pollution correlation characteristic variable prediction substitution model based on numerical simulation and different machine learning algorithms is established for a typical condition framework complex underground water environment system, and an optimal prediction model is preferably selected from the models. Reliable prediction of soil and groundwater environment states, evolution processes and value domain sections of related influences in different application scenes can be accurately and efficiently realized on the premise of satisfying a given condition framework, and the accuracy, the efficiency and the practicability of evaluation and pre-judgment of complex groundwater pollution spatio-temporal evolution states, pollutant spatio-temporal distribution and the like are remarkably improved. Powerful support is provided for water and soil environment system protection and resource safety utilization, and especially for rapid prediction and early warning of soil and groundwater pollution and reliable risk management and control.

Description

Technical Field

[0001] The present invention belongs to the technical field of soil-groundwater pollution prevention and control and risk management, and specifically relates to a method and device for constructing a general prediction model for groundwater pollution based on ANN and SVM. Background Art

[0002] At present, in the monitoring and risk assessment of groundwater pollution, the application of numerical simulation systems to evaluate and quantify groundwater pollution status and environmental risks is a commonly used research method. However, accurate numerical simulation requires comprehensive consideration of many influencing factors and complex evolution processes, which is time-consuming and highly professional. In addition, there is a large uncertainty in the modeling and prediction of complex medium groundwater systems. Uncertainty assessment using methods such as Monte Carlo requires hundreds or thousands of numerical simulation tests, but the computational cost is huge or even unacceptable, and the efficiency is low. Therefore, there is an urgent need for a practical, efficient and simple method to identify the value range segment of groundwater pollution-related characteristic variables. Summary of the invention

[0003] The purpose of the present invention is to provide a method and device for constructing a general prediction model for groundwater pollution based on ANN and SVM, so as to solve the technical problems in the prior art of large uncertainty, high cost and low efficiency in predicting environmental variables of groundwater pollution.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] In a first aspect, the present invention provides a method for constructing a general prediction model for groundwater pollution based on ANN and SVM, comprising:

[0006] Step A: Determine the conceptual model of the spatiotemporal evolution of groundwater pollution under typical conditions and obtain the corresponding typical feature set, i.e., the typical condition framework; select the characteristic variables associated with the spatiotemporal evolution of groundwater pollution; obtain the mathematical control equations for the migration and transformation of groundwater pollutants; and obtain the numerical simulation system for the spatiotemporal evolution of groundwater pollution;

[0007] Step B: Determine the input parameter set V = [v1, v2, ..., v k ], select the range of each model input parameter to form the model input parameter range set U = [u1, u2, ..., u k ], where all the parameters considered and the range of variation of each parameter are determined;

[0008] Step C: Select the change level of each model input parameter, obtain the basic typical scenario sampling of the spatiotemporal evolution of groundwater pollution under a given condition framework based on the orthogonal test method, and obtain the corresponding different model input parameter value combinations of each typical scenario, i.e., parameter combinations;

[0009] Step D: Based on the numerical simulation system of spatiotemporal evolution of groundwater pollution, obtain the simulated values ​​of associated characteristic variables corresponding to each basic typical scenario, and collect all parameter combinations and the corresponding simulated values ​​of associated characteristic variables to form a basic training set for machine learning, which is also a basic data set for statistical modeling;

[0010] Step E: Based on the preset scenario sampling number and the input parameter variable, a preset number of scenario samplings is obtained to obtain a data enhancement scenario parameter combination, and a data enhancement scenario simulation value of the associated characteristic variable corresponding to each scenario is obtained through the numerical simulation system, and the data enhancement scenario parameter combination and the data enhancement scenario simulation value are collected as a machine learning enhancement data set, which is also a statistical modeling enhancement data set;

[0011] Step F: respectively use the machine learning basic training set and the statistical modeling basic data set as the machine learning training set and the statistical modeling basic data set, and respectively use the machine learning enhanced data set and the statistical modeling enhanced data set as the machine learning test set and the statistical modeling verification set; or use the machine learning basic training set and the machine learning enhanced data set as the machine learning training set, the statistical modeling basic data set and the statistical modeling enhanced data set as the statistical modeling data set, in which case, according to step E, the machine learning enhanced data set is obtained, which is also the statistical modeling enhanced data set, and then the machine learning test set and the statistical modeling verification set are obtained;

[0012] Step G: establishing and optimizing a multivariate regression statistical prediction model of the associated characteristic variables and the model input parameter variables according to the statistical modeling data set;

[0013] Step H: obtaining a prediction error value data set of associated characteristic variables through the multivariate regression statistical prediction model and the statistical modeling verification set; based on the prediction error value data set, using a statistical analysis method, determining the upper and lower limits of the distribution interval of the prediction error random variable corresponding to a given preset confidence;

[0014] Step I: using a preset data processing method on the associated characteristic variables, the model input parameter variables and each data set;

[0015] Step J: Based on the obtained simulated values ​​of all associated feature variables corresponding to the basic machine learning training set and the basic machine learning test set, obtain their variation interval [a, b], and use M partition node value (dividing point) sets [y1, y2, …, y M ] is further divided into M+1 sub-intervals, namely: <y1、[y1,y2)、...、[y M-1 ,y M ), ≥y M ;

[0016] Step K: Based on the divided sub - intervals and the simulated values of the associated feature variables of the basic training set and the simulated values of the associated feature variables of the basic test set, obtain the output labels of each scenario of the basic training set and the corresponding output labels of each scenario of the basic test set;

[0017] Step L: Based on the machine - learning basic training set and the machine - learning basic test set, apply the ANN machine - learning modeling principle and the optimization modeling method to obtain a preset number of optimized ANN prediction models. Based on the basic test set and the corresponding output labels of each scenario, obtain the test accuracy of the partition prediction model, and then obtain the optimized ANN prediction model with the highest possible test accuracy of the partition prediction model;

[0018] Step M: Based on the preset scenario sampling number and the model input parameter variables, obtain a preset number of scenario samplings, obtain the data - enhanced scenario parameter combinations. Through the numerical simulation system, obtain the data - enhanced scenario simulation values of the associated feature variables corresponding to each scenario. Determine the prediction confidence interval of the associated feature variables according to the data - enhanced scenario simulation values and the upper and lower limits of the distribution interval of the prediction error random variable corresponding to the given preset confidence level; Based on the model input parameter variables, obtain a preset number of scenario parameter combinations and the prediction confidence interval, and then merge them to obtain the statistical - analysis enhanced modeling data set;

[0019] Step N: Based on the machine - learning basic training set and the corresponding output labels of each scenario, and the machine - learning basic test set and the corresponding output labels of each scenario, and supplement the two - end intervals by applying the statistical - analysis enhanced data set and its corresponding output labels, including <y1 and ≥y M , obtain the SVM machine - learning data set, apply the SVM machine - learning modeling principle and the optimization modeling method, optimize the SVM partition prediction model, and obtain the accuracy of the optimized SVM partition prediction model;

[0020] Step O: Compare the accuracy of the ANN partition prediction model with the accuracy of the SVM partition prediction model, select the machine - learning partition prediction model with higher accuracy, and determine whether the selected machine - learning partition prediction model meets the preset accuracy target. If so, obtain the final machine - learning partition prediction model that meets the model accuracy; If not, compare the accuracy of the ANN partition prediction model with the accuracy of the SVM partition prediction model, select the machine - learning partition prediction model with higher accuracy, and determine whether the selected machine - learning partition prediction model meets the preset accuracy target. If so, obtain the final machine - learning partition prediction model that meets the model accuracy; If not, select any one or any combination of the operations of increasing the basic typical scenario sampling number, the preset scenario sampling number, reducing the preset model accuracy, and reducing the partition node values, execute, and go back to Step E to Step N, complete the corresponding steps until the final machine - learning partition prediction model that meets the model accuracy is obtained;

[0021] Step P: For a given specific application scenario, obtain the specific values ​​of the specific model input parameters, determine whether the specific conditions meet the requirements of the given typical condition framework, and obtain the reliable interval of the predicted values ​​of the associated characteristic variables based on the final machine learning partition prediction model that meets the model accuracy.

[0022] Furthermore, in step A, the conceptual model for determining the spatiotemporal evolution of groundwater pollution under typical conditions specifically includes: atmospheric precipitation infiltration and recharge characteristics, vadose zone characteristics, saturated zone characteristics, pollution source characteristics, groundwater pollution prevention, control and remediation characteristics, source and sink characteristics of injection wells, surface water and groundwater interaction characteristics, groundwater pollutant migration and transformation process types, boundary condition characteristics and initial condition characteristics.

[0023] Furthermore, in step A, the associated characteristic variables 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, stable pollution plume front distance from the pollution source, stable pollution plume average concentration and pollution plume dissipation time; as well as 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, rock-air interface pollutant flux and NAPL pollutant spatiotemporal distribution.

[0024] Furthermore, in step B, the parameter set includes hydraulic conductivity, adsorption parameters, diffusion parameters, and degradation parameters.

[0025] Furthermore, the step E specifically includes: selecting the number of random sampling, determining or selecting the random distribution type and corresponding distribution parameters applicable to each model input parameter variable, completing the random number generation of the model input parameter variable and thus forming a new parameter combination scenario.

[0026] Furthermore, in step I, the preset data processing method includes: retaining the value as it is, increasing the value by the same multiple, decreasing the value by the same multiple, logarithmic conversion of the value, unifying the positive and negative values ​​of the value, and normalizing the value, and any one or any combination of the methods is adopted for processing.

[0027] Furthermore, in step L, the application of ANN machine learning modeling principles and optimization modeling methods includes ANN machine learning method optimization, machine learning modeling optimization method optimization and modeling associated parameter optimization; the ANN machine learning method includes: BP neural network and other machine learning methods including ANN.

[0028] Furthermore, in step L, the optimized ANN prediction model with the highest possible partition prediction model test accuracy is obtained, which includes sorting the ANN prediction models in descending order according to the partition prediction model test accuracy, and the first ANN prediction model that meets the ANN modeling optimization goal is the target model.

[0029] Furthermore, in step N, the optimization modeling method in the application of the SVM machine learning modeling principle and the optimization modeling method includes: optimization of the SVM machine learning model modeling optimization method and optimization of the modeling kernel function type and its associated parameters.

[0030] In a second aspect, the present invention provides a device for constructing a general prediction model for groundwater pollution based on ANN and SVM, comprising:

[0031] The module for constructing the typical condition framework of the groundwater pollution conceptual model is used to determine the conceptual model of the spatiotemporal evolution of groundwater pollution under typical conditions and obtain the corresponding typical feature set, i.e., the typical condition framework;

[0032] The scenario construction and numerical simulation module obtains basic typical scenario sampling of the spatiotemporal evolution of groundwater pollution under a given condition framework based on the orthogonal test method, generates corresponding different model input parameter value combinations for each typical scenario, i.e., parameter combinations, and obtains the simulation values ​​of the corresponding associated characteristic variables of each scenario through the numerical simulation system; at the same time, based on the preset scenario sampling number and input parameter variables, a preset number of scenario samplings are obtained to obtain the machine learning test set and statistical modeling verification set;

[0033] The statistical model building and analysis module builds statistical models and conducts prediction error analysis based on the machine learning basic training set and its corresponding output labels for each scenario, as well as the application of statistical analysis to enhance the data set and its corresponding output labels;

[0034] The machine learning model construction and analysis application module is used to apply machine learning modeling principles and optimization modeling methods, establish and optimize machine learning prediction models, and give prediction results for actual application scenarios.

[0035] Based on the above technical solution, the embodiments of the present invention can at least produce the following technical effects:

[0036] The method for constructing a general groundwater pollution prediction model based on ANN and SVM provided by the present invention selects associated characteristic variables related to the spatiotemporal evolution of groundwater pollution, establishes and optimizes a groundwater pollution associated characteristic variable prediction alternative model based on numerical simulation and different machine learning algorithms for a typical conditional framework complex groundwater environment system, and can accurately and efficiently realize reliable prediction of the value range segments of the soil and groundwater environmental status, evolution process and related impacts in different application scenarios under the premise of meeting a given conditional framework, significantly improve the accuracy, efficiency and practicality of the evaluation and prediction of the spatiotemporal evolution state of complex groundwater pollution and the spatiotemporal distribution of pollutants, and provide strong support for the protection of soil and water environment systems and the safe utilization of resources, especially for the rapid prediction, early warning and reliable risk control of soil and groundwater pollution. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying creative work.

[0038] Figure 1 is a schematic structural diagram of an embodiment of the present invention;

[0039] Figure 2 is the DMPS numerical simulation result in the 54th group of representative scenarios of Example 1 of the present invention;

[0040] Figure 3 It is the T-DMPS numerical simulation result in the representative scenario of Example 54 of the present invention. DETAILED DESCRIPTION

[0041] The technical solutions in the embodiments of the present invention will be described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. In addition, the technical solutions between the various embodiments can be combined with each other, but they must be based on the ability of ordinary technicians in this field to implement them. When the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0042] This embodiment provides a method and device for constructing a general prediction model for groundwater pollution based on ANN (MLB) and SVM. Figure 1-3 Describe them together.

[0043] 1. Selection of associated feature variables

[0044] Firstly, the groundwater system was analyzed according to the profile of different typical study areas, the range of each parameter was determined, and the orthogonal experimental design was used to set the model parameters at different levels. According to the experimental design rules, 54 representative scenarios were selected, and then numerical simulation was performed using GMS software. Each simulation scenario can obtain two characteristic factors of the spatiotemporal evolution of groundwater pollution under the natural attenuation conditions, namely, the stable distance DMPS of the pollution plume and the stable time T-DMPS of the stable pollution plume. After the ln logarithmic processing of the 54 groups of orthogonal experimental numerical simulation results, the multivariate regression model was used for statistics, and the BP and SVM machine learning algorithms were used to establish and optimize the general prediction model of machine learning. Combined with the actual conditions, the specific values ​​of the model input parameters were obtained. The established general prediction model was used to predict the sections where the stable time and migration distance of the groundwater pollution plume are located.

[0045] 2. Generalization of the spatiotemporal evolution of natural attenuation of groundwater pollution

[0046] By referring to the on-site conditions in the study area, a conceptual model of groundwater pollution migration and transformation was established, such as Figure 1 As shown. The groundwater environmental system is generalized from top to bottom into the vadose zone, the phreatic aquifer, the aquitard, the confined aquifer and the impermeable bottom plate. The study area is L in length, B in width and Hz in thickness. The corresponding grid division standard is adopted on the basis of meeting the accuracy of numerical simulation calculation. It is assumed that each layer in the groundwater system is homogeneous and isotropic, and a fixed head boundary is set on the east and west sides, with a specific head difference between the boundaries. There is a constant pollution source in the phreatic layer. The pollutants are easily degradable dissolved organic matter represented by benzene series. The organic pollutants in groundwater decay naturally, and the migration and transformation process of pollutants includes convection, hydrodynamic dispersion, adsorption and degradation.

[0047] 3. Obtain the mathematical control equations for the migration and transformation of groundwater pollutants and obtain the temporal and spatial evolution data of groundwater pollution.

[0048] Value Simulation System

[0049] (1) Establishment of groundwater seepage mathematical model (control equation):

[0050]

[0051] Where: Ω is the seepage area; t is time (T); x, y, z are coordinates (L); K x , K y , K z is the penetration parameter along the x, y, and z directions (LT -1 );ω is the source-sink term, which represents the volume of groundwater flowing out or inflowing from a unit volume of aquifer per unit time (L3 T -1 ), μ is the water supply degree of the aquifer (dimensionless), and H(x, y, z, t) is the groundwater level elevation (L).

[0052] (2) Mathematical model of organic pollutant solute transport (control equation):

[0053] Taking into account convection, diffusion, adsorption and biodegradation, the mathematical control equation for the migration of typical organic pollutant solutes is as follows:

[0054]

[0055] Where: C k is the concentration of pollutants (BTEX) in groundwater (ML -3 );q s is the flow rate into the aquifer per unit volume per unit time, representing the source and sink terms (T -1 );θ is the porosity (dimensionless); D ij is the hydrodynamic dispersion coefficient (L 2 T -1 ); ν i is the groundwater seepage velocity (LT -1 ).

[0056] 4. Determine the parameters and their range of variation, and design typical scenarios through orthogonal experiments

[0057] The natural attenuation of organic pollutants in groundwater involves many parameters. We use scenario analysis to explore them and simulate different results by changing the model parameter assignments. A set of model assignments corresponds to a scenario. Orthogonal experiments, as a statistical method, are often used to solve multi-factor and multi-level experimental design problems. The main idea is to understand the overall experimental situation by analyzing a part of representative experimental results. Using regular orthogonal array design, representative level combinations are selected from all factor and level combinations, which has the advantages of high efficiency, rapidity and economic feasibility.

[0058] A total of 25 parameters that can affect the natural attenuation of groundwater pollution plumes were selected in this orthogonal experimental design. Due to the large number of parameters, according to the design rules of orthogonal experiments, the L54 (2^1, 3^24) orthogonal array design experimental scheme was adopted. Except for the weak permeability effective porosity, which was set at 2 levels, all other parameters were set at 3 levels, generating a total of 54 groups of orthogonal experiments corresponding to 54 simulation scenarios. Each parameter setting is taken from high to low according to its possible actual distribution range. Table 1 lists the different value levels of the 25 parameters that may be involved in the groundwater pollution simulation model.

[0059] Table 1 Different value levels of each parameter

[0060]

[0061] Note: This model selects 25 parameters that may affect the numerical simulation experiment of groundwater pollution, including characteristic parameters related to hydrogeology, hydrochemistry and pollution sources. h is the water head difference between upstream and downstream boundaries; ΔH v is the head difference between the unconfined aquifer and the confined aquifer; R is the precipitation recharge; Cs is the pollution source degradation coefficient; C0 is the pollution source concentration; A is the pollution source area; r is the ratio of the pollution source thickness to the unconfined aquifer thickness; P is the effective porosity; M is the average thickness of the aquifer; Kw is the permeability coefficient; Ka is the adsorption coefficient; Kd is the degradation coefficient; D is the dispersion.

[0062] 5. Obtain numerical simulation results of typical scenarios based on orthogonal test scenarios and numerical simulation systems

[0063] In this study, GMS software was used to numerically simulate 54 representative groundwater pollution natural attenuation simulation scenarios generated according to the orthogonal experimental design results. According to the numerical simulation results, two types of data, the distance of maximum plume spreading (DMPS) of the pollution plume in the groundwater and the time to reach the maximum diffusion distance (T-DMPS) of the pollution plume in each simulation experiment, were selected as the characteristic variables associated with the spatiotemporal evolution of pollution for each representative simulation scenario. Among them, DMPS is calculated by counting the number of grid points in the phreatic aquifer where the concentration exceeds 0.01 mg / L when the pollution plume reaches the farthest distance. T-DMPS is the time required for DMPS. The simulation results of 54 representative scenarios are summarized as follows: Figure 2 and Figure 3 shown.

[0064] 6. According to the range of variation of the parameter variables, the scenario sampling number is set to 20, 20 scenario samplings of the model input parameter variables are obtained, and the data enhancement scenario parameter combination is obtained. The data enhancement scenario simulation value of the associated characteristic variable corresponding to each scenario is obtained through the numerical simulation system.

[0065] 7. Establish a multivariate regression statistical prediction model of associated characteristic variables based on the results of scenario numerical simulation

[0066] Based on 54 sets of orthogonal test scenario parameter combinations and 20 sets of data enhanced sampling scenario parameter combinations and their numerical simulation results, the two pollution plume spatiotemporal evolution associated characteristic variables DMPS and T-DMPS were used as model dependent variables (Y), and full factor model parameters (X) were selected to establish multivariate regression statistical peers. Multivariate regression analysis was used to establish statistical prediction models for DMPS and T-DMPS, and these multivariate regression statistical models were used to predict the scenario values ​​of associated characteristic variables of spatiotemporal evolution of groundwater pollution under natural attenuation conditions.

[0067] After logarithmic processing of the data, the statistical model of the temporal and spatial evolution characteristic factors of groundwater pollution under natural attenuation conditions is as follows:

[0068] lnY=lnρ+α1lnΔH h +α2lnΔH v +α3lnP1+α4lnΔM1+…+α 25 lnr

[0069] After conversion, we get:

[0070]

[0071] Among them, Y represents the spatiotemporal evolution characteristics of the groundwater pollution plume, including DMPS and T-DMPS.

[0072] The DMPS (abbreviated as D) multiple regression statistical model with all model factors selected as independent variables X is expressed as follows:

[0073]

[0074] R 2 =0.939

[0075] The T-DMPS (abbreviated as T) multiple regression statistical model with all model factors selected as independent variables X is expressed as follows:

[0076]

[0077] R 2 =0.909

[0078] Based on the results of variance analysis, a full factorial multivariate regression model was constructed, and its R 2 The value reaches above 0.9. This indicates that there is a strong correlation between the temporal and spatial evolution associated characteristic variables of the groundwater contamination plume under natural attenuation conditions and the model parameters.

[0079] 8. Random sampling to obtain the prediction error of the statistical model and determine the prediction confidence interval of the statistical prediction model

[0080] Given 50 random sampling numbers, 50 scenario samples are obtained based on the model input parameters, and the data-enhanced scenario parameter combination is obtained. The data-enhanced scenario simulation values ​​of the associated characteristic variables corresponding to each scenario are obtained through the numerical simulation system. The predicted values ​​of the statistical model based on the data-enhanced scenario simulation values ​​are used to obtain the upper and lower limits of the distribution interval of the prediction error random variable corresponding to the 5% confidence level, and then the prediction confidence interval of the associated characteristic variables is determined.

[0081] 9. Random sampling to obtain statistical analysis based on SVM modeling to enhance the data set

[0082] Given 1000 random sampling numbers, 1000 scenario samples are obtained based on the model input parameters, and data-enhanced scenario parameter combinations are obtained. Based on the statistical prediction model, the data-enhanced scenario simulation values ​​of the associated characteristic variables corresponding to each scenario are obtained; based on the 1000 scenario parameter combinations and the corresponding associated characteristic variable prediction confidence intervals, the SVM modeling statistical analysis enhanced data set is obtained.

[0083] 10. Construction of groundwater pollution prediction model based on ANN and SVM

[0084] The prediction model was constructed using the full-layer neural network (MLB) in ANN, and 25 influencing factors of natural attenuation in groundwater pollution were used as input parameters of the prediction model, and DMPS and T-DMPS were used as output indicators.

[0085] For the construction of ANN prediction model and SVM prediction model, different value range segments are defined for each associated characteristic variable of DMPS and T-DMPS, different labels are set, and corresponding labels are assigned to the segments where the simulation values ​​are located for different scenarios. Among them, for the two end segments, SVM modeling statistical analysis is superimposed to strengthen the data set.

[0086] The prediction accuracy of both types of machine learning models is low and cannot meet the requirement of prediction accuracy above 0.8.

[0087] 11. The model prediction accuracy is low. Increase random sampling to obtain machine learning modeling reinforcement data set and establish the corresponding machine learning prediction model

[0088] The prediction accuracy of the machine learning model is low, and more sampling scenarios need to be added.

[0089] Given 300 random sampling numbers, 300 scenario samples are obtained according to the model input parameter variables, and a data-enhanced scenario parameter combination is obtained. The data-enhanced scenario simulation value of the associated characteristic variable corresponding to each scenario is obtained through a numerical simulation system; based on the 300 scenario parameter combinations and the corresponding associated characteristic variable simulation values, a machine learning modeling enhanced data set is obtained;

[0090] When the number of samples is increased, the classification results of the SVM prediction model and the neural network prediction model are shown in Tables 2 to 5 below.

[0091] Table 2 SVM model training and testing results

[0092]

[0093] Table 3 SVM model training and testing results

[0094]

[0095] Table 4 SVM model training and testing results

[0096]

[0097] Table 5. MLB neural network model training and test results

[0098]

[0099]

[0100] 12. Increase random sampling again to obtain machine learning modeling reinforcement data set (to 600 groups) and establish the corresponding machine learning prediction model

[0101] Based on 300 sets of data, the amount of data was increased to 600 sets to train the model.

[0102] When the number of samples is increased again, the classification results of the SVM prediction model and the neural network prediction model are shown in Tables 6 to 9 below.

[0103] Table 6 SVM model training and testing results

[0104]

[0105] Table 7 SVM model training and test results

[0106]

[0107] Table 8 SVM model training and test results

[0108]

[0109] Table 9 MLB neural network model training test results

[0110]

[0111]

[0112] 13. Increase random sampling again to obtain machine learning modeling reinforcement data set (to 1200 groups), establish the corresponding machine learning prediction model, and select the machine learning prediction model

[0113] Based on 600 sets of data, the amount of data was increased to 1,200 sets for model training.

[0114] When the number of samples is increased again, the classification results of the SVM prediction model and the neural network prediction model are shown in Tables 10 to 13 below.

[0115] Table 10 SVM model training and test results

[0116]

[0117] Table 11 SVM model training and test results

[0118]

[0119] Table 12 SVM model training and test results

[0120]

[0121] Table 13 MLB neural network model training test results

[0122]

[0123]

[0124] According to the above research results, by comparing the prediction accuracy of the SVM-based model and the ANN-based model, the preferred results of the machine learning prediction model are as follows: the SVM-based prediction model is preferred for the stabilization time of the pollution plume; and the MLB neural network prediction model is preferred for the stable migration distance of the pollution plume.

[0125] 14. Specific scenario application

[0126] The specific application case conditions are as follows: Taking a chemical cluster area in the Yellow River alluvial plain as a typical research area, a conceptual model that can construct different representative simulation scenarios is established. The average annual rainfall in the area is about 600mm, and the average annual evaporation is about 1650mm. The groundwater in the park is mainly the fourth type of loose rock pore water. The groundwater flows from west to east. The overall terrain is flat, and the average annual groundwater level change does not exceed 2m. The distribution range of pollutants exceeding the standard in this area is wide and there are many types, mainly organic indicators such as benzene and toluene.

[0127] Based on the conditions of the given specific application scenario, it is determined that the actual conditions are consistent with the given typical feature set, and that the actual parameters fall within the variation range of the modeling parameters. According to the specific values ​​of the parameters input into the specific model, the obtained preferred machine learning zoning prediction model is applied to predict the specific partition where the characteristic variables related to the specific application scenario are located (the stable time interval is 10-15 years, and the stable distance is 500-600m), which may lead to pollution of downstream sensitive protection receptors. This has achieved reliable prediction and evaluation of the spatiotemporal evolution of the groundwater environment and environmental impacts.

[0128] The above shows and describes the basic principles and main features of the present invention and the advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of the present invention to be protected is defined by the attached claims and their equivalents.

Claims

1. A method for constructing a general prediction model for groundwater pollution based on ANN and SVM, characterized in that: include: Step A: Determine the conceptual model of the spatiotemporal evolution of groundwater pollution under typical conditions and obtain the corresponding typical feature set, i.e., the typical condition framework; select the characteristic variables associated with the spatiotemporal evolution of groundwater pollution; obtain the mathematical control equations for the migration and transformation of groundwater pollutants; and obtain the numerical simulation system for the spatiotemporal evolution of groundwater pollution; Step B: Determine the input parameter set V = [v1, v2, ..., v k ], select the range of each model input parameter to form the model input parameter range set U = [u1, u2, ..., u k ], where all the parameters considered and the range of variation of each parameter are determined; Step C: Select the change level of each model input parameter, obtain the basic typical scenario sampling of the spatiotemporal evolution of groundwater pollution under a given condition framework based on the orthogonal test method, and obtain the corresponding different model input parameter value combinations of each typical scenario, i.e., parameter combinations; Step D: Based on the numerical simulation system of spatiotemporal evolution of groundwater pollution, obtain the simulated values ​​of associated characteristic variables corresponding to each basic typical scenario, and collect all parameter combinations and the corresponding simulated values ​​of associated characteristic variables to form a basic training set for machine learning, which is also a basic data set for statistical modeling; Step E: Based on the preset scenario sampling number and the input parameter variable, a preset number of scenario samplings is obtained to obtain a data enhancement scenario parameter combination, and a data enhancement scenario simulation value of the associated characteristic variable corresponding to each scenario is obtained through the numerical simulation system, and the data enhancement scenario parameter combination and the data enhancement scenario simulation value are collected as a machine learning enhancement data set, which is also a statistical modeling enhancement data set; Step F: respectively use the machine learning basic training set and the statistical modeling basic data set as the machine learning training set and the statistical modeling basic data set, and respectively use the machine learning enhanced data set and the statistical modeling enhanced data set as the machine learning test set and the statistical modeling verification set; or use the machine learning basic training set and the machine learning enhanced data set as the machine learning training set, the statistical modeling basic data set and the statistical modeling enhanced data set as the statistical modeling data set, in which case, according to step E, the machine learning enhanced data set is obtained, which is also the statistical modeling enhanced data set, and then the machine learning test set and the statistical modeling verification set are obtained; Step G: establishing and optimizing a multivariate regression statistical prediction model of the associated characteristic variables and the model input parameter variables according to the statistical modeling data set; Step H: obtaining a prediction error value data set of associated characteristic variables through the multivariate regression statistical prediction model and the statistical modeling verification set; based on the prediction error value data set, using a statistical analysis method, determining the upper and lower limits of the distribution interval of the prediction error random variable corresponding to a given preset confidence; Step I: using a preset data processing method on the associated characteristic variables, the model input parameter variables and each data set; Step J: Based on the obtained simulated values ​​of all associated feature variables corresponding to the basic machine learning training set and the basic machine learning test set, obtain their variation interval [a, b], and use M partition node value sets [y1, y2, …, y M ] is further divided into M+1 sub-intervals, namely: <y1、[y1,y2)、...、[y M-1 ,y M ), ≥y M ; Step K: Obtain output labels for each scenario in the basic training set and output labels corresponding to each scenario in the basic test set according to the divided sub-intervals and the simulated values ​​of the characteristic variables associated with the basic training set and the simulated values ​​of the characteristic variables associated with the basic test set; Step L: Based on the machine learning basic training set and the machine learning basic test set, apply the ANN machine learning modeling principle and optimization modeling method to obtain a preset number of optimized ANN prediction models, and based on the basic test set and the output labels corresponding to each scenario, obtain the partition prediction model test accuracy, and then obtain the optimized ANN prediction model with the highest possible partition prediction model test accuracy; Step M: obtaining a preset number of scenario samples based on a preset scenario sampling number and the model input parameter variable, obtaining a data-enhanced scenario parameter combination, obtaining a data-enhanced scenario simulation value of an associated characteristic variable corresponding to each scenario through the numerical simulation system, and determining the associated characteristic variable prediction confidence interval according to the data-enhanced scenario simulation value and the upper and lower limits of the prediction error random variable distribution interval corresponding to the given preset confidence; Based on the model input parameter variables, a preset number of scenario parameter combinations and the prediction confidence interval are obtained and then combined to obtain a statistical analysis enhanced modeling data set; Step N: Based on the machine learning basic training set and its output labels corresponding to each scenario, the machine learning basic test set and its output labels corresponding to each scenario, and the application statistical analysis to strengthen the data set and its corresponding output labels to supplement both ends of the interval, including <y1 and ≥y M , obtain the SVM machine learning data set, apply the SVM machine learning modeling principle and the optimized modeling method, optimize the SVM partition prediction model, and obtain the accuracy of the optimized SVM partition prediction model; Step O: Compare the accuracy of the ANN partition prediction model with the accuracy of the SVM partition prediction model, select the machine learning partition prediction model with greater accuracy, and determine whether the selected machine learning partition prediction model meets the preset accuracy target. If so, obtain the final machine learning partition prediction model that meets the model accuracy; if not, select any one or any combination of operations of increasing the number of basic typical scenario sampling, the number of preset scenario sampling, reducing the preset model accuracy, and reducing the partition node value, and return to step E to step N to complete the corresponding steps until the final machine learning partition prediction model that meets the model accuracy is obtained; Step P: For a given specific application scenario, obtain the specific values ​​of the specific model input parameters, determine whether the specific conditions meet the requirements of the given typical condition framework, and obtain the reliable interval of the predicted values ​​of the associated characteristic variables based on the final machine learning partition prediction model that meets the model accuracy.

2. The construction method according to claim 1, characterized in that: In step A, the conceptual model for the spatiotemporal evolution of groundwater pollution under typical conditions is specifically determined, including: atmospheric precipitation infiltration and recharge characteristics, vadose zone characteristics, saturated zone characteristics, pollution source characteristics, groundwater pollution prevention, control and remediation characteristics, pumping and injection well source and sink characteristics, surface water and groundwater interaction characteristics, groundwater pollutant migration and transformation process types, boundary condition characteristics and initial condition characteristics.

3. The construction method according to claim 1, characterized in that: In step A, the associated characteristic variables 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, stable pollution plume front distance from the pollution source, stable pollution plume average concentration and pollution plume dissipation time; as well as 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, rock-air interface pollutant flux and NAPL pollutant spatiotemporal distribution.

4. The construction method according to claim 1, characterized in that: The step E specifically includes: selecting the number of random samples, determining or selecting the random distribution type and corresponding distribution parameters applicable to each model input parameter variable, completing the random number generation of the model input parameter variable and thus forming a new parameter combination scenario.

5. The construction method according to claim 1, characterized in that: In step I, the preset data processing method includes: retaining the value as it is, increasing the value by the same multiple, decreasing the value by the same multiple, logarithmic conversion of the value, unifying the positive and negative values ​​of the value and normalizing the value, and any one or any combination of the methods is adopted for processing.

6. The construction method according to claim 1, characterized in that: In step L, the application of ANN machine learning modeling principles and optimization modeling methods includes ANN machine learning method optimization, machine learning modeling optimization method optimization and modeling related parameter optimization; The ANN machine learning method includes: BP neural network and other machine learning methods including ANN.

7. The construction method according to claim 1, characterized in that: In step L, the optimized ANN prediction model with the highest possible partition prediction model test accuracy is obtained, which includes sorting the ANN prediction models in descending order according to the partition prediction model test accuracy, and the first ANN prediction model that meets the ANN modeling optimization goal is the target model.

8. The construction method according to claim 1, characterized in that: In step N, the optimization modeling method in the application of SVM machine learning modeling principle and optimization modeling method includes: optimization of SVM machine learning model modeling optimization method and optimization of modeling kernel function type and its associated parameters.

9. A device for constructing a general prediction model for groundwater pollution based on ANN and SVM, used to implement the construction method according to any one of claims 1 to 8, characterized in that: include: The module for constructing the typical condition framework of the groundwater pollution conceptual model is used to determine the conceptual model of the spatiotemporal evolution of groundwater pollution under typical conditions and obtain the corresponding typical feature set, i.e., the typical condition framework; The scenario construction and numerical simulation module obtains basic typical scenario sampling of the spatiotemporal evolution of groundwater pollution under a given condition framework based on the orthogonal test method, generates corresponding different model input parameter value combinations for each typical scenario, i.e., parameter combinations, and obtains the simulation values ​​of the corresponding associated characteristic variables of each scenario through the numerical simulation system; at the same time, based on the preset scenario sampling number and input parameter variables, a preset number of scenario samplings are obtained to obtain the machine learning test set and statistical modeling verification set; The statistical model building and analysis module builds statistical models and conducts prediction error analysis based on the machine learning basic training set and its corresponding output labels for each scenario, as well as the application of statistical analysis to enhance the data set and its corresponding output labels; The machine learning model construction and analysis application module is used to apply machine learning modeling principles and optimization modeling methods, establish and optimize machine learning prediction models, and give prediction results for actual application scenarios.

Images