A sampling method for constructing a surrogate model in dynamic optimization of contaminated groundwater extraction and treatment

The pumping time in contaminated groundwater extraction treatment is optimized through the optimal Latin hypercube sampling and Rosenblatt transformation method, and a high-precision alternative model is built, which solves the problems of pumping time optimization and sampling uniformity, reducing the management cost and calculation load.

CN117932854BActive Publication Date: 2024-11-15CHINA UNIV OF MINING & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310036470.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-01-10
Publication Date
2024-11-15
Estimated Expiration
2043-01-10

AI Technical Summary

Technical Problem

During the process of extracting and treating polluted groundwater, it is difficult for the existing technology to effectively optimize the pumping time of the pumping well, resulting in an increase in the cost of treatment. The existing alternative model sampling methods cannot ensure that the sampling is evenly distributed in the parameter value space, affecting the calculation accuracy.

Method used

The optimal Latin hypercube sampling method and Rosenblatt transformation are used to use the pumping start and end time of the pumping well as the input variables. The initial sample set is generated through these methods, and the numerical simulation model is trained and tested to build a high-precision alternative. Model.

Benefits of technology

The logical reality of the pumping start time is less than or equal to the end time, ensuring the uniform distribution of samples in the parameter value space, improving the accuracy of the alternative model, and reducing the calculation load and governance cost.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117932854B_ABST
    Figure CN117932854B_ABST
Patent Text Reader

Abstract

The present invention belongs to the field of environmental protection and water resources, and discloses a sampling method for constructing a substitution model in the dynamic optimization of the extraction and treatment of contaminated groundwater, and the steps are as follows: Step 1: Establish a numerical simulation model of groundwater flow and solute migration; Step 2: Use the pumping rate, pumping start time and end time of the candidate pumping well as input variables of the substitution model; Step 3: Perform Rosenblatt transformation on the pumping start time and end time in Step 2; Step 4: Substitute the input sample set in Step 3 into the simulation model of Step 1; Step 5: Repeat Step 2, Step 3 and Step 4 to obtain a test sample set of the substitution model; Step 6: Use the training sample set in Step 4 to construct a substitution model of the numerical model, and perform a test. This method solves the problem that the start time of pumping is greater than the end time when the optimal Latin hypercube sampling method is used for sample sampling in the process of constructing the substitution model, and can ensure the uniform distribution of sample points.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to the technical field of environmental protection and water resources, and in particular to a sampling method for constructing a substitution model in dynamic optimization of extraction and treatment of contaminated groundwater. Background Art

[0002] Groundwater pollution is characterized by concealment, delayed discovery, and great difficulty in treatment. Groundwater extraction and treatment technology is widely used in the treatment of polluted groundwater due to its wide range of applications, short treatment cycle, and simple treatment equipment. In the implementation of this technology, the number, location, pumping volume, and pumping time of the pumping wells largely determine the effectiveness and economy of the treatment project, and also become the key to the application of this technology in the treatment of polluted groundwater.

[0003] In the simulation optimization problem of contaminated groundwater extraction and treatment, current research mostly takes the pumping time of the pumping well as a single variable, that is, the start and end time of pumping for all pumping wells are the same. However, in the actual process of contaminated groundwater extraction and treatment, the pollution plume will continue to move downstream as the treatment time continues. At this time, if the upstream pumping wells in the pollution plume continue to pump water, it will inevitably increase the treatment cost; or when the pollution plume has not yet reached the downstream pumping well, the downstream pumping wells start pumping water at the same time as the upstream pumping wells, which will also increase the treatment cost. Therefore, in the process of optimizing the groundwater extraction and treatment plan, the start and end time of pumping of each candidate pumping well are taken as two independent optimization variables, and the pumping time of each well is dynamically adjusted, which is of great significance to improving the efficiency of contaminated groundwater extraction and treatment and reducing the treatment cost.

[0004] At present, many scholars have coupled groundwater simulation models with optimization algorithms to establish simulation optimization models to optimize extraction treatment schemes. However, in the process of optimizing the extraction treatment scheme for contaminated groundwater, it is necessary to call the numerical simulation model of groundwater flow and solute transport many times, and the computational cost is very high. The substitute model is a substitute for the input-output response relationship of the simulation model. It is functionally close to the simulation model and can effectively reduce the amount of calculation. Therefore, constructing a substitute model for the numerical model has become an effective way to overcome the computational load.

[0005] The degree of approximation of the surrogate model to the numerical simulation model mainly depends on the modeling method and sampling method of the surrogate model. Common sampling methods include Monte Carlo sampling, Latin hypercube sampling, optimal Latin hypercube sampling, uniform design sampling, etc. These sampling methods generally require that the relationship between the parameter variables is independent. In the optimization process of contaminated groundwater extraction and treatment, although the start time and end time of pumping for each pumping well are taken as two variables, in the real world, the two are not completely independent and have a close correlation. It is necessary to ensure that the start time of pumping for each well is less than or equal to the end time. Therefore, how to improve the sample sampling method, so that the start time of pumping for each pumping well is less than or equal to the end time, and ensure that the samples are evenly distributed in the parameter value space, is a key technical issue in realizing the use of surrogate models for contaminated groundwater extraction and treatment. Summary of the invention

[0006] The purpose of the present invention is to provide a sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment, and to establish a numerical simulation model of groundwater flow and solute migration according to the hydrogeological conditions and groundwater pollution conditions of the study area; the pumping rate, pumping start time and end time of the candidate pumping wells are respectively used as input variables of the substitution model and a value range is given, and the optimal Latin hypercube sampling method is used to extract the initial sample set of each variable (at this time, it is temporarily assumed that the pumping start time and end time of each well are independent); the Rosenblatt transformation is applied to the pumping start time and end time of each candidate pumping well in the initial sample set to obtain the input sample set, so that the pumping start time is less than or equal to the end time, and the samples are guaranteed to be evenly distributed in the parameter value space, thereby improving the substitution accuracy of the substitution model.

[0007] To achieve the above-mentioned purpose, the present invention provides the following technical solution: a sampling method for constructing a substitution model in dynamic optimization of contaminated groundwater extraction and treatment, comprising:

[0008] Step 1: Establish a numerical simulation model of groundwater flow and solute transport based on the hydrogeological conditions and groundwater pollution status of the study area, and construct a simulation optimization model for groundwater extraction and treatment;

[0009] Step 2: The pumping rate, pumping start time and end time of the alternative pumping well are used as input variables of the substitution model and given a value range, and the optimal Latin hypercube sampling method is used to extract the initial sample set of each variable;

[0010] Step 3: Perform Rosenblatt transformation on the pumping start and end time of each candidate pumping well in the initial sample set in step 2 to obtain the input sample set of the alternative model;

[0011] Step 4: Bring the input sample set in step 3 into the groundwater flow and solute transport numerical simulation model constructed in step 1 to obtain the output sample set of groundwater level and solute concentration at each monitoring station, and use the input-output sample set of the numerical simulation model as the training sample set of the alternative model;

[0012] Step 5: Repeat the methods of steps 2, 3 and 4 to obtain the test sample set of the alternative model;

[0013] Step 6: Use the training sample set formed in step 4 to construct the water level substitution model and water quality substitution model of the numerical model respectively, and use the test sample set formed in step 5 to test the accuracy of the substitution model.

[0014] Preferably, step 1: establishing a numerical simulation model of groundwater flow and solute transport according to the hydrogeological conditions and groundwater pollution conditions of the study area, and constructing a simulation optimization model for groundwater extraction and treatment;

[0015] Step 2: The pumping rate, pumping start time and end time of the alternative pumping well are used as input variables of the substitution model and given a value range, and the optimal Latin hypercube sampling method is used to extract the initial sample set of each variable;

[0016] Step 3: Perform Rosenblatt transformation on the pumping start and end time of each candidate pumping well in the initial sample set in step 2 to obtain the input sample set of the alternative model;

[0017] Step 4: Bring the input sample set in step 3 into the groundwater flow and solute transport numerical simulation model constructed in step 1 to obtain the output sample set of groundwater level and solute concentration at each monitoring station, and use the input-output sample set of the numerical simulation model as the training sample set of the alternative model;

[0018] Step 5: Repeat the methods of steps 2, 3 and 4 to obtain the test sample set of the alternative model;

[0019] Step 6: Use the training sample set formed in step 4 to construct the water level substitution model and water quality substitution model of the numerical model respectively, and use the test sample set formed in step 5 to test the accuracy of the substitution model.

[0020] Preferably, in step 1, the internationally used Groundwater Modeling System (GMS) or MODFLOW and MT3DMS modules in Visual MODFLOW are used to simulate the groundwater head and solute transport process, and the groundwater extraction treatment optimization model includes an objective function and constraints.

[0021] Objective function:

[0022]

[0023] Constraints:

[0024] x i =0,1(i=1,2,…,N)

[0025] Δt i =t i2 -t i1 , and T1≤t i1 ≤t i2 ≤T1+T

[0026] When x i =1,0 i ≤Q max ,Δt i >0, and t i1 and t i2 When x is an integer i = 0, Q i =0,Δt i =0

[0027] h j ≥h min ,j=1,2,…,N,…,N+M

[0028] C j ≤C0,j=1,2,…,N,…,N+M

[0029] In the formula, F is the cost of treatment, RMB; k1 is the cost of installation of a single well, RMB / well; k2 is the cost of extracting and treating unit volume of groundwater, RMB / m 3 ; T1 is the time when pollution is discovered and treatment begins, days; T is the time limit for treatment, days; Q max Indicates the maximum pumping rate of the pumping well, m 3 / d; the pumping rate of N candidate wells is Q i (i=1,2,…,N); the start time of pumping and the end time of pumping are respectively t i1 ,t i2 (i=1,2,…,N),Δt i =t i2 -t i1 represents the pumping time of alternative pumping well i; x i =1 means the i-th candidate well is installed and used as a pumping well, x i =0 means that the i-th candidate well is not installed; h j represents the simulated water level of well j during the treatment period, where j = 1, 2, ..., N, N + 1, ..., N + M represents the well number, N represents the number of alternative pumping wells, and M represents the number of special monitoring wells; h min ​is the groundwater pumping ban level, m; C j represents the concentration of pollutant Z, mg / L, at the end of the treatment for well j at time ; C0 represents the treatment target concentration, mg / L.

[0030] Preferably, in step 2, the pumping rates, start times, and end times of the alternative pumping wells are determined according to the water-lifting devices configured for each pumping well, and the initial sample size is at least 10 times the number of input variables.

[0031] Preferably, the alternative model construction methods in step 3 mainly include the Kriging method, support vector regression method, kernel extreme learning machine method, machine learning method, etc.

[0032] The Rosenblatt transformation process includes:

[0033] The start time of pumping for the pumping well is X, and the end time is Y. The value ranges of both are D, and the two-dimensional random variable (X, Y) follows a uniform distribution on the region D. The region D is a triangular region enclosed by the y-axis, the line y = 1, and the line y = x, as Figure 1 shown. The joint probability density function of (X, Y) is

[0034] The two marginal probability density functions of (X, Y) are respectively and

[0035] When x < 0, f(x, y) = 0, so f X (x) = 0;

[0036] When x > 1, f(x, y) = 0, so f X (x) = 0;

[0037] When 0 ≤ x ≤ 1,

[0038] In summary, Similarly,

[0039] The Rosenblatt transformation can transform a set of correlated non-normal random variables into a set of equivalent independent normal random variables. Using a similar idea, let

[0040]

[0041] Also, because when 0 < y ≤ 1, f Y (y) > 0, so

[0042] Since U is defined by the distribution function,

[0043] When u<0, P(U≤u)=0;

[0044] When u>1, P(U≤u)=1;

[0045] When 0≤u≤1,

[0046] Similarly, V is defined by the distribution function, so

[0047] When v<0, P(V≤v)=0;

[0048] When v>1, P(V≤v)=1;

[0049] When 0≤v≤1,

[0050] Thus we get

[0051]

[0052]

[0053] It can be seen that U and V obey the uniform distribution on [0,1].

[0054] When 0≤u≤1 and 0≤v≤1, P(U≤u)=u, P(V≤v)=v;

[0055]

[0056] It can be seen from this that U and V are independent of each other.

[0057] In summary, U and V both obey a uniform distribution on [0,1] and are independent of each other.

[0058] When 0≤x≤y≤1, we have roll out

[0059] In order to sample as evenly as possible in area D and ensure that x≤y, the optimal Latin hypercube sampling method is first used to obtain U and V in the square area [0,1] 2 A random and uniform sample set Φ on the The sample set Ψ on the area D is obtained, and the sample set Ψ obtained in this way can ensure the randomness and uniformity on the area D.

[0060] Preferably, the test sample set in step 5 has the same variable value range as the training sample set, and is also obtained through optimal Latin hypercube sampling and Rosenblatt transformation, and the number of samples is at least one fifth of the number of training samples.

[0061] Preferably, the test index of the accuracy of the alternative model in step 6 mainly includes the determination coefficient (R 2 ), mean absolute error (MAE) and root mean square error (RMSE), the formulas are as follows:

[0062] ①Determination coefficient:

[0063] ②Mean absolute error:

[0064] ③Root mean square error: Where n represents the number of output variables of the numerical model, y i represents the output value of the numerical model, represents the output value of the surrogate model, Represents the mean of the numerical model output values.

[0065] Compared with the prior art, the present invention has the following beneficial effects:

[0066] The present invention provides a sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment. In view of the existing optimization design of contaminated groundwater extraction and treatment, the pumping start and end times of all candidate pumping wells are taken as the same variable, and the mobility problem of the pollution plume is ignored. The present invention proposes a sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment. The pumping start and end times of the pumping wells are taken as two independent variables, and a method combining the optimal Latin hypercube sampling method with the Rosenblatt transformation is adopted to realize the logical reality that the pumping start time is less than or equal to the end time in the sampling process, and the sampling uniformity can be guaranteed, thereby improving the substitution accuracy of the substitution model. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 This is a schematic diagram of the area D of the present invention;

[0068] Figure 2 Schematic diagram of the joint distribution function of (U, V) of the present invention;

[0069] Figure 3 Schematic diagram of the initial sample set Φ of the present invention;

[0070] Figure 4 Schematic diagram of sample set Ψ of the present invention;

[0071] Figure 5 Schematic diagram of sample set II of the present invention;

[0072] Figure 6 It is a schematic diagram of a calculation model of the present invention;

[0073] Figure 7 It is a schematic diagram comparing the output results of the Kriging substitution model of the alternative pumping well of the present invention and the simulation model; wherein a is a schematic diagram for comparing water levels, and b is a schematic diagram for comparing pollutant concentrations. DETAILED DESCRIPTION

[0074] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and specific embodiments.

[0075] The present invention proposes a sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment, comprising:

[0076] Step 1: Establish a numerical simulation model of groundwater flow and solute transport based on the hydrogeological conditions and groundwater pollution status of the study area, and construct a simulation optimization model for groundwater extraction and treatment;

[0077] Step 2: The pumping rate, pumping start time and end time of the candidate pumping wells are used as input variables of the substitution model and given a value range. The optimal Latin hypercube sampling method is used to extract the initial sample set of each variable (at this time, it is temporarily assumed that the start and end time of pumping of each well are independent);

[0078] Step 3: Perform Rosenblatt transformation on the pumping start and end time of each candidate pumping well in the initial sample set in step 2 to obtain the input sample set of the alternative model;

[0079] Step 4: Bring the input sample set in step 3 into the groundwater flow and solute transport numerical simulation model constructed in step 1 to obtain the groundwater level and solute concentration (output sample set) of each monitoring station, and use the input-output sample set of the numerical simulation model as the training sample set of the alternative model;

[0080] Step 5: Repeat the methods of steps 2, 3 and 4 to obtain the test sample set of the alternative model;

[0081] Step 6: Use the training sample set formed in step 4 to construct the water level substitution model and water quality substitution model of the numerical model respectively, and use the test sample set formed in step 5 to test the accuracy of the substitution model.

[0082] The internationally accepted GMS or MODFLOW and MT3DMS modules in Visual MODFLOW are used to simulate the groundwater head and solute transport process.

[0083] Groundwater extraction and treatment optimization model, including objective function and constraints.

[0084] Objective function:

[0085]

[0086] Constraints:

[0087] x i =0,1 (i=1,2,…,N) (2)

[0088] Δt i =t i2 -t i1 , and T1≤t i1 ≤t i2 ≤T1+T

[0089] When x i =1,0 i ≤Q max ,Δt i >0, and t i1 and t i2 is an integer (3)

[0090] When x i = 0, Q i =0, Δt i =0 (4)

[0091] h j ≥h min ,j=1,2,…,N,…,N+M (5)

[0092] C j ≤C0,j=1,2,…,N,…,N+M (6) In the formula, F represents the treatment cost, RMB; k1 represents the installation cost of a single well, RMB / well; k2 represents the cost of extracting and treating a unit volume of groundwater, RMB / m 3 ; T1 is the time when pollution is discovered and treatment begins, days; T is the time limit for treatment, days; Q max Indicates the maximum pumping rate of the pumping well, m 3 / d; the pumping rate of N candidate wells is Q i (i=1,2,…,N); the start time of pumping and the end time of pumping are respectively t i1 ,t i2 (i=1,2,…,N),Δt i =t i2 -t i1 represents the pumping time of alternative pumping well i; x i =1 means that the i-th candidate well is installed and used as a pumping well (in this case, 0 i ≤Q max , Δt i ​​>0),x i = 0 means that the i-th candidate well is not installed (in this case, let Q i =0,t i1 =t i2 =T1); h j represents the simulated water level of well j during the treatment period, where j = 1, 2, ..., N, N + 1, ..., N + M represents the well number, N represents the number of alternative pumping wells, and M represents the number of special monitoring wells; h min is the groundwater extraction prohibition level, m; C j Indicates the time after the treatment of well j is completed is the concentration of pollutant X at time t, mg / L; C0 is the treatment target concentration, mg / L.

[0093] The value ranges of the pumping rate, pumping start time and pumping end time of the alternative pumping wells in step 2 are determined according to the water-lifting devices configured for each pumping well, and the initial sample size is at least 10 times the number of input variables.

[0094] The alternative model construction methods in step 3 mainly include Kriging method, support vector regression method, kernel extreme learning machine method, machine learning method, etc.

[0095] The Rosenblatt transformation process is as follows:

[0096] Assume that the pumping start time of the pumping well is X and the end time is Y, the range of their values ​​is D, and the two-dimensional random variable (X, Y) obeys the uniform distribution in area D. Area D is the triangular area surrounded by the y-axis, the straight line y=1 and the straight line y=x, such as Figure 1 As shown, the joint probability density function of (X,Y) is

[0097] The two marginal probability density functions of (X,Y) are and

[0098] When x<0, f(x,y)=0, so f X (x) = 0;

[0099] When x>1, f(x,y)=0, so f X (x) = 0;

[0100] When 0≤x≤1,

[0101] In summary, Similarly,

[0102] The Rosenblatt transformation can transform a set of correlated non-normal random variables into a set of equivalent independent normal random variables. Using a similar idea, let

[0103]

[0104] Also, since \(f(y)>0\) when \(0 < y\leq1\), then Y (y)>0, so

[0105] Since \(U\) is defined by the distribution function, thus

[0106] When \(u < 0\), \(P(U\leq u)=0\);

[0107] When \(u>1\), \(P(U\leq u)=1\);

[0108] When \(0\leq u\leq1\),

[0109] Similarly, since \(V\) is defined by the distribution function, thus

[0110] When \(v < 0\), \(P(V\leq v)=0\);

[0111] When \(v>1\), \(P(V\leq v)=1\);

[0112] When \(0\leq v\leq1\),

[0113] So we get

[0114]

[0115] From this, it can be seen that \(U\) and \(V\) follow a uniform distribution on \([0,1]\).

[0116] When \(0\leq u\leq1\) and \(0\leq v\leq1\), \(P(U\leq u)=u\), \(P(V\leq v)=v\);

[0117]

[0118] As Figure 2 shown, \(S\) represents the area of region \(D\), and \(S'\) represents the area of the light shaded part within region \(D\).

[0119] From this, it can be seen that \(U\) and \(V\) are independent.

[0120] In summary, both \(U\) and \(V\) follow a uniform distribution on \([0,1]\) and are independent;

[0121] When \(0\leq x\leq y\leq1\), there is It is deduced that

[0122] In order to sample as evenly as possible in area D and ensure that x≤y, the optimal Latin hypercube sampling method is first used to obtain U and V in the square area [0,1] 2 A random and uniform sample set Φ on the The sample set Ψ on the area D is obtained, and the sample set Ψ obtained in this way can ensure the randomness and uniformity on the area D.

[0123] Assume that 100 groups of samples are drawn from area D. First, the optimal Latin hypercube sampling method is used to extract the initial sample set Φ( Figure 3 ), and then after Rosenblatt transformation, we get the sample set Ψ( Figure 4 ). In order to compare with the Monte Carlo sampling method, the Monte Carlo method is used to randomly select 100 groups of samples Π( Figure 5 ). By comparison, we can see that after Rosenblatt transformation, x≤y is achieved, and the sample set Ψ is more evenly distributed than the sample set Π.

[0124] Generally speaking, to achieve A≤x≤y≤B ([A,B] represents the parameter value range), it is only necessary to perform interval transformation on the sample points in the above sample set Ψ. Suppose the coordinates of any sample point in Ψ are (x0, y0), and at the same time, it can be guaranteed that x0≤y0. The specific transformation formula is as follows: x=A+(BA)x0, y=A+(BA)y0.

[0125] The test sample set in step 5 has the same variable value range as the training sample set and is also obtained through optimal Latin hypercube sampling and Rosenblatt transformation. The number of test samples is at least one fifth of the number of training samples.

[0126] The test indicators of the accuracy of the alternative model in step 6 mainly include the determination coefficient (R 2 ), mean absolute error (MAE) and root mean square error (RMSE), the formulas are as follows:

[0127] ①Determination coefficient:

[0128] ②Mean absolute error:

[0129] ③Root mean square error: Where n represents the number of output variables of the numerical model, y i represents the output value of the numerical model, represents the output value of the surrogate model, Represents the mean of the numerical model output values.

[0130] Specific

[0131] Assume that the study area is a rectangular area ( Figure 6 ), 660m long and 360m wide. The target aquifer is a sandy confined aquifer with a thickness of 20m. The western boundary Γ1 and the eastern boundary Γ3 are given head boundaries, of which the eastern head is 25.6m and the western head is 30m. The northern boundary Γ2 and the southern boundary Γ4 are water-repellent boundaries. Under natural conditions, groundwater flow is a two-dimensional homogeneous, isotropic, confined, unsteady flow flowing from west to east.

[0132] The background value of groundwater quality in the study area is good, and the initial concentration of pollutants in the aquifer is zero. One day, pollutant Z was found in the downstream of the study area. According to investigation, the pollutant came from a chemical enterprise upstream. The enterprise injected water into the water well (200m 3 / d) continuously and steadily discharge pollutants into the aquifer. There are 17 wells in the pollution plume, of which 10 wells are used as alternative pumping wells and observation wells, and 7 wells are dedicated monitoring wells. It is known that the installation cost of the pumping well is 5,000 yuan / well, and the cost of extracting and treating pollutants Z is 0.3 yuan / m 3 , the maximum pumping rate of a single well Q max =300m 3 / d. To ensure water supply safety, it is required to use existing wells in the polluted area within 200 days from the date of discovery of pollution, adopt the method of extraction and treatment, and reduce the pollutant concentration to the limit value of 6mg / L at the minimum treatment cost, and the water level cannot be lower than the top plate of the target aquifer during the extraction process.

[0133] Table 1 Known hydrogeological parameters in the study area

[0134]

[0135] (1) Groundwater flow and solute transport model

[0136] A certain time when no pollution occurs is taken as the initial time, at which time t0 = 0. Assuming that the chemical enterprise discharges pollutants from the 5th to the 15th day, water quality testing is carried out on the 600th day, and pollutant Z is found (its safe drinking limit is 6 mg / L). According to the coordinate system and hydrogeological conditions of the study area, a groundwater flow numerical model is established:

[0137]

[0138] Where: K is the permeability coefficient of the aquifer, m / d; H is the water level of the confined aquifer, m; w is the source and sink term; S s is the water storage rate, m -1 ;Ω is the range of the simulation area; H0(x,y) is the initial water level, m; H1 is the known water head on a type of boundary, m; Γ1, Γ3 are given water head boundaries; Γ2, Γ4 are zero flow boundaries; is the permeability coefficient on the boundary normal vector, m / d; is the external normal direction of the second type boundary.

[0139] Based on the groundwater flow numerical model, a groundwater solute transport numerical model is established. The boundaries of the simulation area can be generalized as follows: Γ1 is the zero concentration boundary, Γ3 is the convection diffusion flux boundary, and Γ2 and Γ4 are regarded as zero diffusion flux boundaries. The solute transport model of groundwater in the study area is:

[0140]

[0141] Where: D x , D y is the component of the hydrodynamic dispersion coefficient in the x and y directions, m 2 / d;v x 、v y are the seepage velocity of groundwater in the x and y directions, m / d; n is the porosity of the aquifer medium, dimensionless; c is the mass concentration of the pollutant, mg / L; Q inj is the amount of liquid injected into the aquifer, m 3 / d;C inj is the pollutant concentration entering the aquifer with the liquid, mg / L; f(x,y,t) represents the mass of solute passing through the actual water flow section per unit time under the action of hydrodynamic diffusion alone.

[0142] The established groundwater flow and solute transport simulation model was calculated and solved using the MODFLOW and MT3DMS modules in the GMS software. To improve the simulation accuracy, the study area was divided into a square finite difference grid with 120 rows and 220 columns, and the basic cell side length was 3m.

[0143] (2) Simulation Optimization Model

[0144] The objective function of the extraction and treatment scheme optimization model is to minimize the treatment cost, which mainly includes the installation cost of the pumping well, the groundwater extraction and treatment cost, etc. Constrained by the treatment cost and treatment cycle, the optimization variables of the optimization model mainly include the number, location, pumping rate of the pumping wells, and the start and end time of the pumping of each alternative pumping well.

[0145] The mathematical expressions of the optimization model are shown in equations (1)-(6).

[0146] In the process of solving the optimization model, the groundwater flow and solute transport simulation model needs to be called repeatedly. In order to reduce the computational load, an alternative model of the groundwater flow and solute transport model is constructed.

[0147] Therefore, equations (5) and (6)

[0148] h≥hmin,j=1,2,…,N,…,N+M

[0149] C j ≤C0,j=1,2,…,N,…,N+M

[0150] is replaced by an alternative model, expressed as follows:

[0151]

[0152] In the formula, represents the water level of the alternative model output at the end of pumping in well j, m; It represents the concentration of pollutant Z output by the alternative model at the end of pumping for well j, mg / L.

[0153] (3) Construction of alternative models

[0154] The pumping rate Q of 10 alternative pumping wells i (i=1,2,…,10) and the pumping start time t of each well i1 and end time t i2 (i=1,2,…,10) are used as the input variables of the substitution model, and the water levels and pollutant concentrations of the alternative pumping wells and special monitoring wells (a total of 17 wells) are used as output variables. The Kriging method is used to build the substitution model.

[0155] Since it is not determined whether the candidate pumping well is a pumping well before sampling, the optimal Latin hypercube sampling method and Rosenblatt transformation are combined to select the pumping rate Q i (i=1,2,…,10) value range [0,300] and the pumping start time t of each well i1 and end time t i2 1000 groups of samples are drawn from the range [600,800] with uniform distribution (ensuring t i1 ≤t i2 ), as the input value of the Kriging surrogate model.

[0156] 1000 groups of input samples were substituted into the GMS software to obtain the daily water level and pollutant concentration values ​​of 10 alternative pumping wells and 7 special monitoring wells within [600d, 801d] as the output values ​​of the Kriging substitution model. The 1000 groups of input values ​​and output values ​​were used as the training sample set, and the Kriging substitution model of water level and the Kriging substitution model of pollutant concentration of the numerical simulation model were constructed by the Kriging method. Then, using the same sampling process, 200 groups of samples were re-extracted from the pumping rate range and treatment time range of the 10 alternative pumping wells as the input values ​​of the test samples, and substituted into the GMS software and Kriging substitution model respectively, and the daily water level and pollutant concentration values ​​of 17 wells within [600d, 801d] were obtained, and the 200 groups of input values ​​and output values ​​were used as the test sample set.

[0157] Taking the 1# alternative pumping well as an example, the water level value h output by the simulation model of the test sample i,out As the horizontal axis, the water level value output by the Kriging substitution model As the vertical axis, a comparison chart of the output results of the groundwater level simulation model and the Kriging substitution model is plotted ( Figure 7 a); Similarly, the pollutant concentration value C output by the numerical simulation model i,out As the horizontal axis, the pollutant concentration value output by the Kriging substitution model As the vertical axis, a comparison chart of the pollutant concentration simulation model and the Kriging substitution model output is plotted ( Figure 7 b). Using the coefficient of determination (R 2 ), mean absolute error (MAE) and root mean square error (RMSE) were used to test the accuracy of the alternative model. The results show that the method combining the optimal Latin hypercube sampling method and the Rosenblatt transformation can achieve the pumping start time being less than or equal to the end time, and the constructed Kriging alternative model has high accuracy and can be used as an approximate alternative to the numerical simulation model.

[0158] Table 2 Determination coefficient and mean absolute error of Kriging substitution model for 17 monitoring wells

[0159]

[0160]

[0161] The above-mentioned specific embodiments are only several preferred embodiments of the present invention. Based on the technical solutions of the present invention and the relevant inspirations of the above-mentioned embodiments, those skilled in the art can make various alternative improvements and combinations to the above-mentioned specific embodiments.

Claims

1. A sampling method for constructing a substitution model in dynamic optimization of contaminated groundwater extraction and treatment, characterized in that: include: Step 1: According to the hydrogeological conditions and groundwater pollution status of the study area, a numerical simulation model of groundwater flow and solute transport is established, and a simulation optimization model for groundwater extraction and treatment is constructed; Step 2: The pumping rate, pumping start time and end time of the alternative pumping well are used as input variables of the substitution model and given a value range, and the optimal Latin hypercube sampling method is used to extract the initial sample set of each variable; Step 3: Perform Rosenblatt transformation on the pumping start and end time of each candidate pumping well in the initial sample set in step 2 to obtain the input sample set of the alternative model; Step 4: Bring the input sample set in step 3 into the groundwater flow and solute transport numerical simulation model constructed in step 1 to obtain the output sample set of groundwater level and solute concentration at each monitoring station, and use the input-output sample set of the numerical simulation model as the training sample set of the alternative model; Step 5: Repeat the methods of steps 2, 3 and 4 to obtain the test sample set of the alternative model; Step 6: Use the training sample set formed in step 4 to respectively construct the water level substitution model and water quality substitution model of the numerical model, and use the test sample set formed in step 5 to test the accuracy of the substitution model; Alternative model building methods in step 3 include Kriging method, support vector regression method and kernel extreme learning machine method; The Rosenblatt transformation process includes: The pumping start time of the pumping well is , the end time is , the range of their values ​​is D, and the two-dimensional random variable It follows a uniform distribution over region D, which is given by Axis, Linear and straight line The enclosed triangular area, The joint probability density function of ; The two marginal probability density functions are and ; when hour, ,thereby ; when hour, ,thereby ; when hour, ; In summary, , Similarly, ; The Rosenblatt transformation can transform a set of correlated non-normal random variables into a set of equivalent independent normal random variables. ; Because when hour, ,thereby , ; because Defined by the distribution function, when hour, ; when hour, ; when hour, ; Similarly, Defined by the distribution function, when hour, ; when hour, ; when hour, ; Thus we get , ; , ; It is determined from this , It obeys the uniform distribution on [0,1]; when and hour, , ; ; in, represents the area of ​​region D, represents the area of ​​light shade in region D, thus determining , Independent of each other; In summary, , All obey the uniform distribution on [0,1] and are independent of each other; when Sometimes, there are ,roll out ; In order to sample as evenly as possible in area D and ensure , first use the optimal Latin hypercube sampling method to get In the square area [0,1] 2 A random and uniform sample set on , and then use the above functional relationship Get the sample set on area D , so the sample set obtained The randomness and uniformity in area D are guaranteed.

2. The sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment according to claim 1, characterized in that: In step 1, the MODFLOW and MT3DMS modules in the internationally used Groundwater Modeling System or Visual MODFLOW are used to simulate the groundwater head and solute transport process. The groundwater extraction and treatment optimization model includes objective functions and constraints.

3. The sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment according to claim 2, characterized in that: Objective function: ; Constraints: ; In the formula, represents the governance cost, yuan; It indicates the installation cost of a single well, Yuan / well; Indicates the cost of extracting and treating unit volume of groundwater, yuan / m 3 ; The time to discover pollution and start treatment, days; Indicates the governance time limit, days; represents the maximum pumping rate of the pumping well, m 3 / d; Pumping rate of the selected well It indicates that, ; The start time of pumping and the end time of pumping are recorded as , ,in, ; Indicates alternative pumping well Pumping time; Indicates The alternative well is installed and used as a pumping well. Indicates The eye alternative well is not installed; Indicates well The simulated water level during the treatment period is: represents the pound sign, where represents the number of alternative pumping wells, Indicates the number of dedicated monitoring wells; is the groundwater extraction prohibition level, m; Indicates well After the governance is over The concentration of pollutant Z at time , mg / L; Indicates the target concentration for treatment, mg / L.

4. The sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment according to claim 1, characterized in that: The value ranges of the pumping rate, pumping start time and pumping end time of the alternative pumping wells in step 2 are determined according to the water-lifting devices configured for each pumping well, and the initial sample size is at least 10 times the number of input variables.

5. The sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment according to claim 1, characterized in that: The test sample set in step 5 has the same variable value range as the training sample set and is also obtained through optimal Latin hypercube sampling and Rosenblatt transformation. The number of test samples is at least one fifth of the number of training samples.

6. The sampling method for constructing a substitution model in the dynamic optimization of contaminated groundwater extraction and treatment according to claim 1, characterized in that: The test indicators of the accuracy of the alternative model in step 6 mainly include the determination coefficient , mean absolute error MAE and root mean square error RMSE, the formula is as follows: ①Determination coefficient: ; ②Mean absolute error: ; ③Root mean square error: ; In the formula represents the number of output variables of the numerical model, represents the output value of the numerical model, represents the output value of the surrogate model, Represents the mean of the numerical model output values.