Method for predicting adsorption quantity of pollutants in soil based on combined use of linear programming, optimization algorithm and neural network

Abstract

The invention relates to a method for predicting the adsorption quantity of pollutants in soil based on combined use of linear programming, an optimization algorithm and a neural network, and belongs to the field of ecological environmental protection, and the method comprises the following steps: collecting a soil sample in a research area, and obtaining physicochemical property data of the soil sample; carrying out an indoor adsorption experiment to obtain adsorption quantity data of pollutants under different experiment conditions; calculating mineral quantitative composition of the soil sample by adopting a linear programming algorithm; arranging and dividing a data set, training a neural network model, and obtaining an optimal hyper-parameter by using an optimization algorithm to obtain an optimal prediction model; and applying the neural network model to soil pollutant adsorption quantity prediction. According to the method, it is reasonably assumed that pollutant adsorption capacity is jointly controlled by types and contents of minerals in soil, forms and contents of active iron, soil granularity and environmental conditions, quantitative composition data of the minerals in the soil is acquired by adopting a linear programming algorithm, a neural network model is trained by taking the quantitative composition data as input parameters, and optimal hyper-parameters are acquired by using an optimization algorithm; the prediction precision, generalization ability and prediction ability of the neural network in predicting the soil pollutant adsorption capacity are improved.

Description

Technical Field

[0001] The present invention belongs to the field of ecological environmental protection and can be used to predict the adsorption amount of pollutants in soil media during environmental impact assessment. More specifically, it relates to a method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network. Background Art

[0002] Widespread industrial production activities such as chemical, metallurgical, and mining will release pollutants into the environment, causing soil and groundwater pollution risks. Therefore, it is necessary to carry out environmental impact assessments for related construction projects. Environmental impact prediction is one of the important tasks of environmental impact assessment. The water flow-solute coupled transport model is an important means to predict the migration of pollutants in the soil and groundwater environment, which depends on the accurate determination of the parameters related to the adsorption of pollutants in the soil. The current method mainly obtains relevant parameters by searching references or empirical values. However, the properties of natural soils are variable, and even in a small area, its adsorption performance for pollutants may vary greatly. The parameters given in the literature are obviously unable to accurately describe the adsorption behavior of pollutants in all sites to be evaluated.

[0003] It is a relatively accurate method to calibrate the adsorption parameters of pollutants in soil through indoor adsorption experiments. However, the adsorption of pollutants in soil is affected by many factors, including pollutant concentration, pH value, ionic strength, competing ions, temperature and water-soil ratio. There are too many influencing factors to consider in the experiment, resulting in a huge workload. In addition, indoor adsorption experiments are relatively delicate, which places extremely high demands on the professional skills of technicians. It is difficult for general practitioners in this industry to reach a qualified technical level. As an unconventional experiment, the adsorption parameters of pollutants in soil cannot be obtained through samples sent by general analysis and testing institutions. Therefore, it is time-consuming, costly, cumbersome, and difficult to control the quality of each soil sample to conduct indoor experiments separately. It is difficult to promote it comprehensively. There is an urgent need to develop a method that can accurately predict the adsorption behavior of pollutants in soil, which is easy to use and easy to promote among practitioners.

[0004] Machine learning provides a feasible solution to the above problems. Professionals collect soil samples in a specific area, conduct indoor adsorption experiments on certain pollutants, and train machine learning models based on experimental data. For general practitioners, as long as the site of environmental impact assessment is within this specific area, general physical and chemical property data can be obtained by collecting soil samples for testing, and the machine learning model can be used to predict the adsorption of pollutants by soil samples under specific pollutants and specific experimental conditions, thereby obtaining more accurate adsorption parameters.

[0005] When establishing a machine learning model, the selection of input parameters is a key issue. Existing machine learning models mainly use soil physical and chemical properties and environmental conditions as key factors affecting the amount of pollutant adsorption, among which soil physical and chemical properties mainly refer to soil pH and chemical composition. However, the chemical composition of the soil does not actually directly affect the amount of pollutants adsorbed, because the mineral composition of two soil samples with the same chemical composition may be very different. From the perspective of physical processes, the mineral composition of the soil is the factor that can directly affect the amount of pollutants adsorbed. In addition, the particle size of the soil will also have an important impact on the adsorption of pollutants by affecting the specific surface area. In general, existing machine learning models are mostly separated from the understanding of physical processes and rely solely on data to establish, and there is still room for improvement in the prediction accuracy and generalization ability of pollutant adsorption. Summary of the invention

[0006] To solve the above problems, the present invention proposes a method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network. The specific technical solution is as follows:

[0007] A method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network, comprising the following steps:

[0008] Step 1: Collect soil samples in the study area and obtain the physical and chemical properties data of the soil samples;

[0009] Step 2: Conduct indoor adsorption experiments to obtain adsorption data of pollutants under different experimental conditions;

[0010] Step 3: Calculate the mineral quantitative composition of soil samples using a linear programming algorithm;

[0011] Step 4: Organize and divide the data set, train the neural network model, and use the optimization algorithm to obtain the optimal hyperparameters to obtain the best prediction model;

[0012] Step 5: Use the neural network model to predict the amount of soil pollutant adsorption.

[0013] Furthermore, the physical and chemical property data of the soil sample in step 1 include the following data: qualitative composition of minerals, quantitative composition of chemical components, loss on ignition, iron mineral morphology and particle size distribution data. Specifically:

[0014] S101, collect soil samples at evenly distributed points within the study area, dry the samples naturally in the shade, beat them with a grinding pestle to disperse the united soil particles, pick out the animal and plant residues and keep them as the original soil for later use, take a part of each original sample and grind it with an agate mortar and pass it through a 200-mesh sieve as the ground sample for later use;

[0015] S102, testing the ground sample using XRD, and qualitatively analyzing the mineral composition of the soil sample based on the obtained diffraction results;

[0016] S103, testing the ground sample using XRF to obtain quantitative composition data of chemical components of the soil sample;

[0017] S104, using a muffle furnace to burn the ground sample at 1000°C for 3 hours to determine the loss on ignition of the soil sample;

[0018] The calculation formula for loss on ignition is:

[0019]

[0020] Wherein, LOI is the loss on ignition of the soil sample, M0 is the weight of the porcelain crucible, M1 is the weight of the porcelain crucible after loading the ground sample, and M2 is the weight of the porcelain crucible after ignition.

[0021] S105, conduct iron form extraction experiment on the original soil sample to obtain soil iron form content data; specifically: use 0.5 mol / L formic acid solution to extract carbonate-bound iron, use 0.1 mol / L ascorbic acid solution to extract weakly crystalline iron oxides, use 0.1 mol / L ascorbic acid + 0.2 mol / L ammonium oxalate solution to extract strongly crystalline iron oxides, all solutions are adjusted to pH 3.0 using HCl or NaOH solution, the solid-liquid ratio used for form extraction is 0.2g:25mL, the iron content in the extract is determined by atomic absorption spectrometry, and the solid phase iron form content (mg / g) is 0.125 times the liquid phase iron form concentration (mg / L).

[0022] S106, using a laser particle size analyzer to test the original sample to obtain particle size distribution data of the soil sample.

[0023] Furthermore, in step 2, the experimental conditions are selected from one or more of the pollutant concentration, pH, coexisting ion types and concentrations, temperature, and solid-liquid ratio. The remaining conditions other than the conditions to be predicted should be limited in the solution preparation and subsequent experiments. For example, when the condition to be predicted is the pollutant concentration, pollutant solutions of different concentrations should be prepared, but the pH of the solution, the coexisting ion concentration, and the temperature and solid-liquid ratio used in the experiment should be constant, and the only condition variable of the adsorption experiment is the pollutant concentration. The specific adsorption experiment adopts the following sub-steps:

[0024] S201, selecting predicted experimental conditions according to actual needs, preparing a series of pollutant solutions according to the determined conditions, and taking initial samples of the pollutant solutions;

[0025] S202, weigh a certain amount of original soil sample and mix it with the prepared pollutant solution, and take a sample of the equilibrium solution after shaking for 24 hours;

[0026] S203, measuring the concentration of pollutants in the initial solution and the equilibrium solution, and calculating the adsorption amount of pollutants in the soil.

[0027] The calculation formula for the amount of pollutants adsorbed by soil is:

[0028]

[0029] Where M is the adsorption amount of pollutants in the soil, c0 is the pollutant concentration in the initial solution, and c e is the pollutant concentration of the equilibrium solution, and R is the solid-liquid ratio adopted in the experiment.

[0030] Furthermore, the step 3 specifically includes the following sub-steps:

[0031] S301, according to the molecular formula of the mineral component identified in sub-step S102, convert the molecular formula into the form of mineral standard composition and ignition loss mass fraction according to the element composition and the relative atomic mass of the element;

[0032] S302, based on the quantitative composition data of the chemical components of the soil samples obtained in sub-step S103 and the loss on ignition data of the soil samples obtained in sub-step S104, with the mineral standard composition converted in sub-step S301 as the restriction condition, and the absolute error of the soil chemical components as the condition for judging the optimal solution, use the Linprog function of Matlab software to calculate the mineral quantitative composition of the soil through a linear programming algorithm.

[0033] The linear programming algorithm is as follows:

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] Where n is the number of mineral components contained in the sample, m is the number of chemical components involved in the mineral components, and A mn is the proportion of the mth component in the molecular formula of the nth mineral; W nis the proportion of the nth mineral in the sample; P m is the proportion of the mth chemical component in the sample; α m and β m are the calculated positive and negative errors.

[0043] Furthermore, the step 4 specifically includes the following sub-steps:

[0044] S401, arranging the acquired data into a multi-input single-output data set, wherein the input parameters include soil iron form extraction data, soil mineral quantitative composition data, soil particle size distribution data, and measured data of the conditions to be predicted, and the output is soil pollutant adsorption data;

[0045] S402, dividing the sorted data set into three parts: a training set, a validation set, and a test set, with the ratios of 7:2:1 respectively;

[0046] S403, normalize the training set data and synchronously normalize the validation set data according to the criterion:

[0047] S404, establishing and using training set data to train a neural network model;

[0048] S405, taking the root mean square error of the validation set data on the model as the fitness value, automatically optimizing the hyperparameters of the neural network model through an optimization algorithm to obtain the best neural network model;

[0049] S406, saving the optimal neural network model and the training set normalization criterion.

[0050] Furthermore, in the sub-step S401, the soil iron form extraction data includes four parameters: carbonate iron, weakly crystalline iron, strongly crystalline iron and total iron content, wherein the total iron content is obtained by sub-step S103. The soil particle size distribution data specifically refers to the cumulative frequencies corresponding to the particle sizes of 0.002, 0.01, 0.05 and 0.25 mm, respectively indicating the proportion of clay particles, medium silt particles, coarse silt particles and fine sand particles and below in the soil.

[0051] Furthermore, in the sub-step S403, the formula for data normalization is:

[0052]

[0053] Among them, X norm is the normalized data, X is the data before normalization, and X max is the maximum value in the data set, X min is the minimum value in the data set.

[0054] Furthermore, in the sub-step S405, the root mean square error calculation formula of the model is:

[0055]

[0056] Where RMSE is the root mean square error of the model predictions, n is the number of data points in the validation set, is the mean of the predicted values ​​of the validation set, and is the true value of the validation set.

[0057] Furthermore, the step 5 specifically includes the following sub-steps:

[0058] S501, reading the data to be predicted, and normalizing the data to be predicted according to the neural network and data normalization criteria saved in sub-step S406;

[0059] S502, using the optimal neural network to predict the pollutant adsorption results. If the data to be predicted is changed to the test set data, the generalization ability and prediction ability of the model can be evaluated.

[0060] Furthermore, in the sub-step S502, the indicators for evaluating the generalization ability and prediction ability of the model through the test set include consistency index (IA) and U statistic (UI), and the calculation formula is:

[0061]

[0062]

[0063] Among them, IA is the consistency index of the model prediction. The closer its value is to 1, the stronger the generalization ability of the model is. U1 is the U statistic predicted by the model. The closer its value is to 0, the better the prediction ability of the model is. n is the number of data points in the validation set, is the true value of the test set, is the average of the true values ​​of the test set, and is the predicted value of the test set.

[0064] Based on the physical nature of the adsorption process, the present invention assumes that the amount of adsorption of specific pollutants in the soil is jointly affected by the type and content of minerals in the soil, soil particle size, and environmental conditions. In particular, in addition to the large amount of crystalline minerals, the small amount of iron oxides contained in the soil also has an important influence on the adsorption of pollutants, and its adsorption capacity changes with the form of iron oxides (carbonate iron, weakly crystalline iron, and strongly crystalline iron). Based on this physical assumption, the present invention uses a linear programming algorithm to calculate the quantitative composition of soil minerals, and uses a sequential extraction method to obtain soil iron form content data, thereby forming a training set data for the neural network, and uses an optimization algorithm to automatically optimize hyperparameters during the training process.

[0065] The present invention reasonably assumes that the amount of pollutant adsorption is jointly controlled by the type and content of minerals in the soil, the form and content of active iron, soil particle size and environmental conditions, adopts a linear programming algorithm to obtain the quantitative composition data of minerals in the soil, uses this as input parameters to train a neural network model, and uses an optimization algorithm to obtain the optimal hyperparameters, thereby improving the prediction accuracy, generalization ability and prediction ability of the neural network in predicting the amount of soil pollutant adsorption. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 Flow chart of the method of the present invention.

[0067] Figure 2 Comparison of the true values ​​and predicted values ​​of the training set and the validation set in the embodiment and the root mean square error.

[0068] Figure 3 The predicted results of pollutant adsorption in the examples.

[0069] Figure 4 A comparison diagram of the LP-PSO-BP neural network model in the embodiment and the PSO-BP neural network model and the LP-BP neural network model in the comparative example. DETAILED DESCRIPTION

[0070] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of 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.

[0071] Example

[0072] This embodiment provides a method for predicting the adsorption amount of pollutant arsenic in soil under different concentration conditions based on LP algorithm, particle swarm optimization algorithm (PSO) and BP neural network. Figure 1 As shown, the specific steps include:

[0073] (1) Collect soil samples in the study area and obtain data on the physical and chemical properties of the soil samples.

[0074] In specific implementation, the steps include:

[0075] S101, an area of ​​about 1800km in northern my country 2A total of 80 groups of surface soil samples were collected at evenly distributed points in the basin, with a collection depth of 0-20cm. The soil samples were dried naturally in a cool place, and the soil was beaten with a grinding pestle to disperse the cohesive soil. The animal and plant residues in the soil were picked up and kept as the original soil for later use. A portion of each original sample was ground with an agate mortar and passed through a 200-mesh sieve as a ground sample for later use;

[0076] S102, the ground samples were sent to a qualified institution for XRD spectrum testing. The test was scanned at a speed of 5 degrees per minute within the range of 5 to 90 degrees. The test data were qualitatively analyzed using JADE software. The search results showed that the soil samples contained seven crystal minerals: clinochlore, illite, albite, potassium feldspar, calcite, dolomite, and quartz.

[0077] S103, the ground samples were sent to a qualified institution for XRF spectrum testing, and the content data of K2O, CaO, Na2O, MgO, Al2O3, Fe2O3, and SiO2 in the soil samples were obtained;

[0078] S104, weigh a certain amount of ground sample using a burnt-dried porcelain crucible, burn it at 1000°C for 3 hours using a muffle furnace, weigh the weight after burning, and calculate the loss on ignition of the soil sample according to the following formula:

[0079]

[0080] Wherein, LOI is the loss on ignition of the soil sample, M0 is the weight of the porcelain crucible, M1 is the weight of the porcelain crucible after loading the ground sample, and M2 is the weight of the porcelain crucible after ignition.

[0081] S105, prepare 0.5 mol / L formic acid solution, adjust pH to 3.0 with a small amount of HCl or NaOH solution, weigh 0.2g of soil sample, add 25mL of extracting solution, shake the suspension at 180 rpm for 24h at room temperature using a shaker to extract carbonate iron in the soil, then centrifuge at 4000 rpm for 10 minutes, take the upper clarified sample without disturbing the bottom soil, and use the lower soil for the next extraction. Then, with the same operation steps, use 0.1mol / L ascorbic acid solution (adjust pH=3) and 0.1mol / L ascorbic acid + 0.2mol / L ammonium oxalate solution (adjust pH=3) to extract weak crystalline iron oxides and strong crystalline iron oxides, respectively. Use atomic absorption spectrometer to take the total iron content of the solution, and obtain the corresponding form of iron content (mg / g) in the soil by multiplying the liquid iron form concentration (mg / L) by 0.125;

[0082] S106, sending the soil sample to a qualified institution for testing, and obtaining the particle size distribution data of the soil sample by a laser particle size analyzer.

[0083] (2) Conduct indoor adsorption experiments to obtain adsorption data of pollutants under different experimental conditions.

[0084] In specific implementation, the steps include:

[0085] S201, prepare a sufficient amount of 10mmol / L NaHCO3 buffer solution, and adjust the pH to 7.0 with 6mol / L hydrochloric acid solution. Use high-purity sodium arsenite as the solute and the buffer solution as the solvent to prepare 10 groups of arsenic solutions with different concentrations ranging from 0 to 2mg / L, and collect a small amount of arsenic-containing solution as is for testing. Since the condition to be predicted in this embodiment is the pollutant concentration, the pH of the solution is fixed at 7.0 by using a buffer solution, and the coexisting ion concentration is set to 10mmol / L Na + and HCO3 - ;

[0086] S202, weighing 0.5 g of soil sample into a 30 mL PE bottle, adding 25 mL of the arsenic-containing solution of different concentrations, oscillating and balancing the suspension at room temperature for 24 hours, and then collecting the balanced solution sample for testing using a 0.45 μm water filter;

[0087] S203, using an atomic fluorescence spectrophotometer, the arsenic concentration of the original sample of the arsenic-containing solution and the equilibrium solution sample is measured by hydride generation method, and the adsorption amount of pollutant arsenic in the soil is calculated, and finally 709 sets of valid data are obtained. The calculation formula of the pollutant adsorption amount is as follows:

[0088]

[0089] Where M is the adsorption amount of pollutants in the soil, c0 is the pollutant concentration in the initial solution, and c e is the pollutant concentration of the equilibrium solution, and R is the solid-liquid ratio adopted in the experiment.

[0090] (3) A linear programming algorithm is used to calculate the quantitative mineral composition of soil samples.

[0091] In specific implementation, the steps include:

[0092] S301, according to the 7 crystalline minerals identified in sub-step S102, as well as the organic matter components (SOM) and free iron oxides present in the soil but not identified by XRD, the standard molecular formula of the crystalline minerals is searched from the Water4f.dat database provided by the geochemical software PHREEQC, and the molecular formula is converted into the chemical component form corresponding to sub-steps S103 and S104 in terms of mass percentage. The results are as follows:

[0093] Table 1: Standard element composition of soil minerals

[0094] Chlorite Illite Sodium feldspar Orthoclase Calcite dolomite quartz Iron oxide SOM <![CDATA[K2O]]> 0 7.36 0 16.9 0 0 0 0 0 CaO 0 0 0 0 56.0 30.4 0 0 0 <![CDATA[Na2O]]> 0 0 11.8 0 0 0 0 0 0 MgO 36.3 2.62 0 0 0 21.9 0 0 0 <![CDATA[Al2O3]]> 18.3 30.5 19.4 18.3 0 0 0 0 0 <![CDATA[Fe2O3]]> 0 0 0 0 0 0 0 100 0 <![CDATA[SiO2]]> 32.4 54.8 68.8 64.8 0 0 100 0 0 LOI 13.0 4.69 0 0 44.0 47.7 0 0 100

[0095] S302, based on the quantitative composition data of the chemical components of the soil samples obtained in sub-step S103 and the LOI data of the soil samples obtained in sub-step S104, with the mineral standard composition converted in sub-step S301 as the restriction condition, and the absolute error of the soil chemical components as the condition for judging the optimal solution, the Linprog function in the Matlab 2016 optimization toolbox is used to calculate the mineral quantitative composition of the soil through a linear programming algorithm.

[0096] (4) Organize and divide the data set, train the BP neural network, and use the particle swarm algorithm to optimize the hyperparameters to obtain the best prediction model.

[0097] In specific implementation, the steps include:

[0098] S401, organize the acquired data into a multi-input single-output data set, where the input parameters include soil iron form extraction data, soil mineral quantitative composition data, soil particle size distribution data, and measured data of the conditions to be predicted, and the output is soil pollutant adsorption data. Among them, soil iron form extraction data specifically refers to the content of active iron forms, including three parameters: carbonate iron, weakly crystalline iron, and strongly crystalline iron. Soil mineral quantitative composition data refers to the content data of various crystalline and amorphous components, including clinochlore, illite, albite, potassium feldspar, calcite, dolomite, quartz, and SOM content, a total of 8 parameters. Soil particle size distribution data specifically refers to the cumulative frequency corresponding to particle sizes of 0.002, 0.01, 0.05, and 0.25 mm, including 4 parameters. The conditions to be predicted refer to the equilibrium concentration of pollutants in the solution. A total of 709 groups of valid data were sorted out, each group of data contains 16 input parameters and 1 output parameter.

[0099] S402, randomly dividing the sorted valid data set into three parts: a training set, a validation set, and a test set, wherein there are 500 training sets, 122 validation sets, and 87 test sets;

[0100] S403, normalize the training set data and synchronously normalize the validation set data according to the criterion:

[0101] S404, initializing the parameters of the particle swarm algorithm, wherein the learning factors C1 and C2 are initialized to 4.494, the number of population updates is 50, the population size is 5, the maximum speed is 1.0, the minimum speed is -1.0, the maximum boundary is 1.0, the minimum boundary is -1.0, and the fitness function is set to the root mean square error of the validation set;

[0102] S405, establishing a BP neural network, and using the training set data to train the neural network model;

[0103] S406, optimizing the hyperparameters of the neural network model by particle swarm algorithm to obtain the optimal BP neural network model, such as Figure 2 Shown is the comparison of the true value and predicted value of the training set and the validation set, as well as the root mean square error;

[0104] S407, save the optimal BP neural network and save the training set normalization criterion.

[0105] (5) Use the neural network model to predict the amount of soil pollutant adsorption.

[0106] In specific implementation, the steps include:

[0107] S501, taking the test set as the data to be predicted, normalizing the data to be predicted according to the BP neural network and data normalization criterion saved in sub-step S407;

[0108] S502, using the optimal BP neural network to predict the pollutant adsorption amount of the test set, the prediction result is as follows Figure 3 shown.

[0109] Comparative Example:

[0110] As an example to further demonstrate the beneficial effects of the present invention, two other unimproved neural network models are provided for comparison, including a PSO-BP neural network model that does not use a linear programming algorithm to calculate the quantitative composition of soil minerals as an input parameter, and a LP-BP neural network model that does not use a particle swarm algorithm for hyperparameter optimization. For the PSO-BP neural network model, all operations in step 3 are not performed, and the soil mineral quantitative composition data in the data set described in step S401 is changed to the soil chemical component quantitative composition data, and the other steps are the same as the embodiment. For the LP-BP neural network model, the operations described in substeps S404 and S406 are not performed, and the other steps are the same as the embodiment. The RMSE, IA, and UI of the test set are used to evaluate the generalization ability and prediction performance of the model, and the calculation formula is:

[0111]

[0112]

[0113]

[0114] Among them, RMSE is the root mean square error of the model prediction. The smaller its value, the better the model prediction performance. IA is the consistency index of the model prediction. The closer its value is to 1, the stronger the generalization ability of the model. U1 is the U statistic of the model prediction. The closer its value is to 0, the better the prediction ability of the model. n is the number of data points in the validation set, is the true value of the test set, is the average of the true values ​​of the test set, is the predicted value of the test set, and is the average of the predicted values ​​of the test set.

[0115] Figure 4 The figure shows the evaluation parameters of the LP-PSO-BP neural network model based on the embodiment, as well as the comparative PSO-BP neural network model and the LP-BP neural network model. As can be seen from the figure, the RMSE of the test set prediction of the neural network model based on LP-PSO-BP is 41.4, IA is 0.952, and U1 is 0.036; the RMSE of the PSO-BP neural network model without the linear programming algorithm is 50.2, IA is 0.935, and U1 is 0.052; the RMSE of the LP-BP neural network model without the particle swarm optimization algorithm is 48.1, IA is 0.937, and U1 is 0.041. Using RMSE to indicate the prediction error, the use of the particle swarm optimization algorithm and the linear programming algorithm reduces the prediction error of the BP neural network model by 14% and 17.5%, respectively. Using IA to indicate the generalization ability of the model, the use of the particle swarm optimization algorithm and the linear programming algorithm enhances the generalization ability of the BP neural network model by 1.6% and 1.8%, respectively. With U1 indicating the prediction ability of the model, the use of the particle swarm optimization algorithm and the linear programming algorithm enhances the prediction ability of the BP neural network model by 0.52% and 1.7% respectively. It can be seen that the method of the present invention realizes the prediction of the adsorption amount of soil pollutants under different concentration conditions, and the improvement of the present invention reduces the prediction error of the traditional BP neural network model and improves the generalization ability and prediction ability of the model.

[0116] The above describes in detail the preferred implementation of this patent, but this patent is not limited to the above implementation. Various changes can be made within the knowledge scope of ordinary technicians in this field without departing from the purpose of this patent.

Claims

1. A method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network, characterized in that: The following steps are involved: Step 1: Collect soil samples in the study area and obtain the physical and chemical properties data of the soil samples; Step 2: Conduct indoor adsorption experiments to obtain adsorption data of pollutants under different experimental conditions; Step 3: Calculate the mineral quantitative composition of soil samples using a linear programming algorithm; Step 4: Organize and divide the data set, train the neural network model, and use the optimization algorithm to obtain the optimal hyperparameters to obtain the best prediction model; Step 5: Use the neural network model to predict the amount of soil pollutant adsorption.

2. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 1 is characterized in that: In step 1, the physical and chemical property data of the soil sample include qualitative mineral composition, quantitative chemical component composition, loss on ignition, iron mineral form and particle size distribution data.

3. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 1 is characterized in that: In step 1, the physical and chemical properties of the soil include the morphological content of soil iron minerals, which are extracted using the following method: using 0.5 mol / L formic acid solution to extract carbonate-bound iron, using 0.1 mol / L ascorbic acid solution to extract weakly crystalline iron oxides, and using 0.1 mol / L ascorbic acid + 0.2 mol / L ammonium oxalate solution to extract strongly crystalline iron oxides. All solutions are adjusted to pH 3.0 using HCl or NaOH solution, and the solid-liquid ratio used for morphological extraction is 0.2 g:25 mL.

4. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 1 is characterized in that: In step 2, the experimental conditions are selected from one or more of the following: pollutant concentration, pH, coexisting ion types and concentrations, temperature, and solid-liquid ratio.

5. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 1 is characterized in that: The step 3 includes the following sub-steps: S301, according to the molecular formula of the mineral component identified in step 1, convert the molecular formula into the form of the standard composition of the mineral and the mass fraction of the loss on ignition according to the element composition and the relative atomic mass of the element; S302, based on the quantitative composition data of the chemical components of the soil samples obtained in step 1 and the loss on ignition data of the soil samples obtained in step 1, with the mineral standard composition converted in sub-step S301 as a restriction condition, and with the absolute error sum of the soil chemical components as a condition for determining the optimal solution, the Linprog function of Matlab software is used to calculate the quantitative mineral composition of the soil through a linear programming algorithm; The linear programming algorithm is as follows: Among them, n is the number of mineral components contained in the sample, m is the number of chemical components involved in the mineral components, Amn is the proportion of the mth component in the molecular formula of the nth mineral; Wn is the proportion of the nth mineral in the sample; Pm is the proportion of the mth chemical component in the sample; αm and βm are the positive and negative errors of the calculation.

6. The method for predicting the adsorption amount of pollutants in soil based on linear programming, optimization algorithm and neural network combined use according to claim 1 is characterized in that: The step 4 specifically includes the following sub-steps: S401, arranging the acquired data into a multi-input single-output data set, wherein the input parameters include soil iron form extraction data, soil mineral quantitative composition data, soil particle size distribution data and measured data of the conditions to be predicted, and the output is soil pollutant adsorption data; S402, dividing the sorted data set into three parts: a training set, a validation set, and a test set, with the ratios of 7:2:1 respectively; S403, normalize the training set data and synchronously normalize the validation set data according to the criterion: The formula for data normalization is: Among them, X norm is the normalized data, X is the data before normalization, and X max is the maximum value in the data set, X min is the minimum value in the data set; S404, establishing and using training set data to train a neural network model; S405, taking the root mean square error of the validation set data on the model as the fitness value, automatically optimizing the hyperparameters of the neural network model through an optimization algorithm to obtain the best neural network model; The root mean square error calculation formula of the model is: Where RMSE is the root mean square error of the model prediction, n is the number of data points in the validation set, is the average of the predicted values ​​of the validation set, and is the true value of the validation set; S406, saving the optimal neural network model and the training set normalization criterion.

7. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 6, characterized in that: In the sub-step S401, the soil iron form extraction data includes carbonate iron, weakly crystalline iron, strongly crystalline iron and total iron content, wherein the total iron content is obtained by step 1; the soil particle size distribution data specifically refers to the cumulative frequencies corresponding to particle sizes of 0.002 mm, 0.01 mm, 0.05 mm, and 0.25 mm, respectively indicating the proportions of clay particles, medium silt particles, coarse silt particles, and fine sand particles and below in the soil.

8. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 6, characterized in that: In the sub-step S404, the neural network selects a model of regression problem, such as BP, support vector machine or random forest; in the sub-step S405, the optimization algorithm selects genetic algorithm or particle swarm algorithm.

9. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 6, characterized in that: The step 5 specifically includes the following sub-steps: S501, reading the data to be predicted, and normalizing the data to be predicted according to the neural network and data normalization criteria saved in sub-step S406; S502, using the optimal neural network to predict the pollutant adsorption results. If the data to be predicted is changed to the test set data, the generalization ability and prediction ability of the model can be evaluated.

10. The method for predicting the adsorption amount of pollutants in soil based on the combined use of linear programming, optimization algorithm and neural network according to claim 9, characterized in that: In the sub-step S502, the indicators of the generalization ability and prediction ability of the model evaluated by the test set include the consistency index IA and the U statistic UI, and the calculation formula is: Among them, IA is the consistency index of the model prediction. The closer its value is to 1, the stronger the generalization ability of the model is. U1 is the U statistic predicted by the model. The closer its value is to 0, the better the prediction ability of the model is. n is the number of data points in the validation set, is the true value of the test set, is the average of the true values ​​of the test set, and is the predicted value of the test set.

Images