Membrane pollution pretreatment optimization control method and system based on machine learning

Abstract

The invention discloses a machine learning-based membrane pollution pretreatment optimization control method and system, and the method comprises the steps: constructing a high-dimensional feature data set for Fenton pretreatment prevention and control of membrane pollution, and carrying out the comprehensive pretreatment of data; training and optimizing the candidate machine learning model based on the decision tree; evaluating the trained machine learning model, and selecting the model with the most generalization ability as a Fenton pretreatment control membrane pollution prediction model; based on the prediction model, reversely searching the optimal reagent dosage under a preset environment variable by adopting a genetic algorithm; and carrying out membrane pollution Fenton pretreatment under the corresponding environment variable condition by utilizing the optimal reagent adding amount. According to the technical scheme, the reagent adding amount of Fenton pretreatment with the optimal membrane pollution retarding capacity of different pollutants under different membrane filtration conditions can be optimized and directionally controlled, and the Fenton pretreatment device can be suitable for different types of pollutants and membrane assemblies and has the characteristics of low cost, high efficiency, accuracy, reliability and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of pollution prevention and control prediction, and in particular to an optimization control method for membrane pollution pretreatment based on machine learning and an optimization control system for membrane pollution pretreatment based on machine learning. Background Art

[0002] Membrane separation technology has functions such as separation, concentration, and purification, and has advantages such as high efficiency, small floor area, no phase change, and green and pollution-free, and is widely used in the field of advanced treatment of sewage and wastewater around the world, becoming an important technical means to solve water resource shortage and sewage reuse. However, in the actual application process, a series of problems such as the decline of membrane flux, the deterioration of effluent quality, and the increase of operating costs caused by membrane pollution are the main barriers restricting the development and application of membrane separation technology. In the past, many membrane pollution researchers have conducted a large number of studies on membrane pollution prevention and control from multiple aspects such as pretreatment, membrane modification, and membrane cleaning. Among them, pretreatment technology has become one of the main strategies for membrane pollution prevention and control due to its advantages such as simple operation, low cost, and no side effects.

[0003] Previous studies on pretreatment to mitigate membrane pollution mainly removed pollutants through two technologies: oxidation and coagulation. Traditional oxidation pretreatment technologies mainly controlled membrane pollution by changing the properties of pollutants and causing changes in the interfacial forces of pollutants on the membrane. Common oxidants include KMnO4, H2O2, O3, etc. Coagulation pretreatment, on the other hand, prevented pollutants from entering the membrane pores and blocking membrane pollution by regulating the size of pollutants. Common coagulants include PAC, PFS, PAM, etc. However, the prevention and control effects of membrane pollution by separate oxidation and coagulation pretreatment are limited, and with the shortage of water resources and the improvement of water quality requirements, they can no longer meet the requirements for membrane pollution prevention and control. Using both oxidation and coagulation technologies simultaneously not only increases costs but also may pose a risk of exacerbating membrane pollution. The Fenton technology can generate a large amount of strongly oxidizing ·OH through the action of H2O2 and Fe 2+ and simultaneously generate Fe with coagulation effect 3+ and shows excellent ability to mitigate membrane pollution as a pre-treatment measure before the membrane.

[0004] However, problems such as the harsh environmental conditions, high cost, and secondary pollution of the Fenton reaction have always been the key challenges hindering the application of the Fenton pretreatment technology in the field of membrane pollution mitigation. In previous studies on Fenton pretreatment to control membrane pollution, first, the pH needs to be adjusted to acidic conditions, and only this process requires the addition of a large amount of reagents. And the Fenton pretreatment reaction will also add a large amount of Fe 2+And H2O2 are used to achieve the degradation and mineralization of pollutants. This not only increases the operating cost, but also results in a large amount of iron sludge in the effluent, causing secondary pollution. In fact, the dosage required for membrane fouling prevention and control using the Fenton pretreatment technology for different pollutants is not the same. Some researchers have found that by adding a small amount of Fenton reagent under neutral conditions, a significant reduction in membrane fouling can be achieved. In the actual process, the dosage of Fenton pretreatment is determined solely by experience or selected through traditional single-factor experimental analysis methods. This process not only requires a large amount of manpower and material resources, but also takes a lot of time. Summary of the Invention

[0005] In view of the above problems, the present invention provides a machine learning-based optimization control method and system for membrane fouling pretreatment. Different machine learning models are trained through a high-dimensional feature dataset of multiple variables, and the most generalization-capable model is selected through preset indicators. Combined with the genetic algorithm, the optimal reagent dosage is predicted under certain environmental variable conditions, so as to optimize and directionally control the reagent dosage of the best Fenton pretreatment for reducing the membrane fouling ability of different pollutants under different membrane filtration conditions. This solves the defects of high cost, large investment, and long time-consuming of traditional single-factor experimental analysis, and can be applied to different types of pollutants and membrane modules, with the characteristics of low cost, high efficiency, accuracy, and reliability.

[0006] To achieve the above object, the present invention provides a machine learning-based optimization control method for membrane fouling pretreatment, including:

[0007] Construct a high-dimensional feature dataset for Fenton pretreatment to prevent and control membrane fouling with different environmental variables, and perform comprehensive data preprocessing on the high-dimensional feature dataset;

[0008] Use the high-dimensional feature dataset to train and optimize candidate machine learning models based on decision trees;

[0009] Evaluate the trained machine learning models using preset model evaluation indicators, and select the most generalization-capable machine learning model as the Fenton pretreatment control membrane fouling prediction model;

[0010] Based on the Fenton pretreatment control membrane fouling prediction model, use the genetic algorithm to perform a reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables;

[0011] Use the optimal reagent dosage for Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions.

[0012] In the above technical solution, preferably, the specific process of constructing the high-dimensional feature dataset for Fenton pretreatment to prevent and control membrane fouling with different environmental variables includes:

[0013] Construct a high-dimensional feature dataset of membrane fouling after Fenton pretreatment under different environmental variables, where the environmental variables include temperature, pressure, pH, flow rate, filtration time, pollutant type, pollutant concentration, COD concentration, Fe 2+ dosage, Fe 3+ dosage, H2O2 dosage, the ratio of Fe to H2O2, Fenton reaction time, membrane type, MWCO, initial permeation flux, COD removal rate, and standard permeation flux J / J0;

[0014] The high-dimensional feature dataset includes missing values and outliers, and the feature types include numerical and categorical features.

[0015] In the above technical solution, preferably, the specific process of comprehensively preprocessing the data of the high-dimensional feature dataset includes:

[0016] Perform data preprocessing operations such as data denoising, outlier removal, data normalization, feature selection, and feature transformation on the high-dimensional feature dataset.

[0017] In the above technical solution, preferably, use the high-dimensional feature dataset to train and optimize a candidate machine learning model based on a decision tree. The specific process includes:

[0018] Select LightGBM, XGBoost, and CatBoost models based on decision trees as candidate machine learning models;

[0019] Divide the high-dimensional feature dataset into a training set and a test set according to a ratio of 50% - 90% / 50% - 10%, and ensure that the data in the test set does not appear in the training set;

[0020] Use the training set for five-fold cross-validation, with the standard permeation flux J / J0 as the prediction target, and temperature, pH, flow rate, filtration time, pollutant type, pollutant COD concentration, Fe 2+ dosage, H2O2 dosage, membrane type, MWCO, and initial permeation flux as 11 input features, and train the candidate machine learning models respectively based on the gradient boosting strategy;

[0021] During the training process, use the Optuna optimization framework based on the Bayesian (Bayesian Optimization) algorithm to globally optimize the hyperparameters of the candidate machine learning models to obtain the optimal hyperparameter combination for each model.

[0022] In the above technical solution, preferably, the stratified sampling method is used to divide the high-dimensional feature data set into a training set and a test set, and the division ratio of the training set and the test set is 90% / 10%.

[0023] In the above technical solution, preferably, the trained machine learning model is evaluated using preset model evaluation metrics, and the machine learning model with the best generalization ability is selected as the Fenton pretreatment control membrane fouling prediction model. The specific process includes:

[0024] Using the mean absolute error, root mean square error, and corrected determination coefficient as the preset model evaluation metrics to evaluate each trained machine learning model;

[0025] According to the evaluation results of the mean absolute error, root mean square error, and corrected determination coefficient of each machine learning model, select the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model;

[0026] Among them, the mean absolute error MAE, root mean square error RMSE, and corrected determination coefficient Adj_R 2 The calculation formulas are as follows:

[0027]

[0028] Among them, n is the number of samples in the data set, p is the number of features, and y i is the true standard permeation flux J / J0 of the i-th sample, is the standard permeation flux J / J0 predicted by the model for the i-th sample, is the average standard permeation flux J / J0 of all samples; the lower the MAE and RMSE, and the higher the Adj_R 2 the better the prediction performance of the Fenton pretreatment control membrane fouling prediction model.

[0029] In the above technical solution, preferably, based on the Fenton pretreatment control membrane fouling prediction model, a genetic algorithm is used to inversely search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables. The specific process includes:

[0030] Based on the Fenton pretreatment control membrane fouling prediction model, with the standard permeation flux as the target, the operations of the genetic algorithm are used to inversely search for the optimal Fe 2+ and H2O2 dosages for membrane fouling prevention and control under different preset environmental variables.

[0031] In the above technical solution, preferably, the membrane fouling pretreatment optimization control method based on machine learning further includes:

[0032] Based on the Fenton pretreatment control membrane fouling prediction model, the standard permeation flux under the condition of the optimal reagent dosage is predicted and verified by using the test set, and the optimal reagent dosage is experimentally verified in the membrane filtration system;

[0033] Perform Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions by using the verified optimal reagent dosage.

[0034] In the above technical solution, preferably, if the prediction and verification of the Fenton pretreatment control membrane fouling prediction model by the test set fails or the experimental verification fails, the candidate machine learning model is retrained, optimized and evaluated.

[0035] The present invention also proposes an optimization control system for membrane fouling pretreatment based on machine learning, which applies the optimization control method for membrane fouling pretreatment based on machine learning disclosed in any one of the above technical solutions, including:

[0036] A data set preprocessing module, which is used to construct a high-dimensional feature data set for preventing and controlling membrane fouling by Fenton pretreatment with different environmental variables, and perform comprehensive data preprocessing on the high-dimensional feature data set;

[0037] A candidate model training module, which is used to train and optimize a candidate machine learning model based on a decision tree by using the high-dimensional feature data set;

[0038] An optimal model selection module, which is used to evaluate the trained machine learning model by using a preset model evaluation index, and select the machine learning model with the most generalization ability as the Fenton pretreatment control membrane fouling prediction model;

[0039] An optimal variable prediction module, which is used to reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables based on the Fenton pretreatment control membrane fouling prediction model by using a genetic algorithm;

[0040] An optimal variable implementation module, which is used to perform Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions by using the optimal reagent dosage.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows: Different machine learning models are trained with a high-dimensional feature dataset of multiple variables, and the model with the best generalization ability is selected through preset metrics. Combined with the genetic algorithm, the optimal reagent dosage is predicted under certain environmental variable conditions, so as to optimize the reagent dosage for the best Fenton pretreatment of the membrane fouling mitigation ability of different pollutants under different membrane filtration conditions, solving the defects of high cost, large investment, and long time-consuming of traditional single-factor experimental analysis, and being applicable to different types of pollutants and membrane modules, with the characteristics of low cost, high efficiency, accuracy, and reliability. Description of the Drawings

[0042] Figure 1 It is a schematic flow chart of an optimization control method for membrane fouling pretreatment based on machine learning disclosed in an embodiment of the present invention;

[0043] Figure 2 It is a schematic diagram showing the influence of the division ratio of the training set and the test set on the model prediction performance disclosed in an embodiment of the present invention;

[0044] Figure 3 It is a schematic diagram showing the prediction performance evaluation of three candidate models under the optimal hyperparameter combination disclosed in an embodiment of the present invention;

[0045] Figure 4 It is about the fouling behavior of sodium alginate (SA) on PVDF ultrafiltration membrane under the Fenton pretreatment dosage of 0.15 mM Fe 2+ and 0.09 mM H2O2;

[0046] Figure 5 It is about the fouling behavior of sodium alginate (SA) on PVDF ultrafiltration membrane under the Fenton pretreatment dosage of 0.08 mM Fe 2+ and 0.049 mM H2O2;

[0047] Figure 6 It is about the fouling behavior of sodium alginate (SA) on PVDF ultrafiltration membrane under the Fenton pretreatment dosage of 0.39 mM Fe 2+ and 0.43 mM H2O2;

[0048] Figure 7 It is about the fouling behavior of sodium alginate (SA) on PVDF ultrafiltration membrane under the Fenton pretreatment dosage of 0.47 mM Fe 2+ and 0.15 mM H2O2. Detailed Embodiments

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0050] The following further describes the present invention in detail with reference to the accompanying drawings:

[0051] As Figure 1 shown, a method for optimizing the control of membrane fouling pretreatment based on machine learning provided by the present invention includes:

[0052] Construct a high-dimensional feature dataset for preventing and controlling membrane fouling by Fenton pretreatment with different environmental variables, and perform comprehensive data preprocessing on the high-dimensional feature dataset;

[0053] Use the high-dimensional feature dataset to train and optimize a candidate machine learning model based on a decision tree;

[0054] Use a preset model evaluation index to evaluate the trained machine learning model, and select the machine learning model with the strongest generalization ability as the Fenton pretreatment control membrane fouling prediction model;

[0055] Based on the Fenton pretreatment control membrane fouling prediction model, use a genetic algorithm to perform a reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables;

[0056] Use the optimal reagent dosage to perform Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions.

[0057] In this embodiment, different machine learning models are trained through a high-dimensional feature dataset with multiple variables, and the model with the strongest generalization ability is selected through preset indexes. Combining with a genetic algorithm, the optimal reagent dosage prediction under certain environmental variable conditions is realized, so as to optimize the directional control of the reagent dosage of Fenton pretreatment with the best membrane fouling mitigation ability for different pollutants under different membrane filtration conditions, solve the defects of high cost, large investment, and long time-consuming of traditional single-factor experimental analysis, and can be applied to different types of pollutants and membrane components, effectively improving the prevention and control efficiency of membrane fouling in the actual application process of membrane water treatment technology, and having the characteristics of low cost, high efficiency, accuracy, and reliability.

[0058] Specifically, a high-dimensional feature dataset of membrane fouling after Fenton pretreatment under different multivariate environments is constructed, and a comprehensive processing framework for normalization, feature selection, and transformation is established; the processed Fenton pretreatment-controlled membrane fouling dataset is input into different models for training; the trained Fenton pretreatment-controlled membrane fouling models are evaluated to determine the best model; based on the Fenton pretreatment-controlled membrane fouling model, the genetic algorithm is used to inversely determine the optimal Fe 2+ and H2O2 dosage; the Fenton pretreatment-controlled membrane fouling model is used to predict and verify the membrane fouling behavior under the optimal Fe 2+ and H2O2 dosage.

[0059] Among them, the data source of the high-dimensional feature dataset can adopt relevant research literature on Fenton pretreatment-controlled membrane fouling in the past. All data related to Fenton pretreatment-controlled membrane fouling in the literature pictures are extracted using the "EngagementDigitizer" graphic digitization software. This method requires no investment in any additional experiments and operating costs. Only by collecting membrane fouling data under different experimental conditions based on previous research can the training of the model be completed.

[0060] In the above embodiment, preferably, in the high-dimensional feature dataset of membrane fouling after Fenton pretreatment under different environmental variables, the environmental variables include temperature, pressure, pH, flow rate, filtration time, pollutant type, pollutant concentration, COD concentration, Fe 2+ dosage, Fe 3+ dosage, H2O2 dosage, the ratio of Fe to H2O2, Fenton reaction time, membrane type, MWCO, initial permeation flux, COD removal rate, and standard permeation flux J / J0; where J0 is the initial permeation flux and J is the real-time permeation flux.

[0061] The high-dimensional feature dataset includes missing values and outliers, and the feature types include numerical and categorical features.

[0062] In the above embodiment, preferably, the specific process of data comprehensive pretreatment of the high-dimensional feature dataset includes:

[0063] Data preprocessing operations such as data denoising, outlier removal, data normalization, feature selection, and feature transformation are performed on the high-dimensional feature dataset.

[0064] Among them, for the data comprehensive pretreatment of the high-dimensional feature dataset, a comprehensive processing framework can be pre-constructed, and data denoising, outlier removal, data normalization, feature selection, and feature transformation and other processing operations are respectively performed on the data in the high-dimensional feature dataset.

[0065] In the above embodiments, preferably, a high-dimensional feature dataset is used to train and optimize a candidate machine learning model based on a decision tree. The specific process includes:

[0066] Select models with efficient feature selection capabilities and powerful non-linear fitting capabilities, such as LightGBM, XGBoost, and CatBoost models based on decision trees, as candidate machine learning models;

[0067] The high-dimensional feature dataset is divided into a training set and a test set according to a ratio of 50% - 90% / 50% - 10%, and it is ensured that the data in the test set does not appear in the training set;

[0068] Using the training set with five-fold cross-validation, with the standard permeate flux J / J0 as the prediction target, and temperature, pH, flow rate, filtration time, pollutant type, pollutant COD concentration, Fe 2+ Dosage, H2O2 dosage, membrane type, MWCO, and initial permeate flux as 11 input features, which are respectively input into the above candidate machine learning models and trained based on the gradient boosting strategy;

[0069] During the training process, the Optuna optimization framework based on the Bayesian algorithm is used to globally optimize the hyperparameters of the candidate machine learning models to obtain the optimal hyperparameter combination for each model.

[0070] Among them, using five-fold cross-validation to train the model and optimize the hyperparameters can prevent the model from overfitting. Through the Bayesian inference mechanism, the optimal hyperparameter combination of the model can be accurately captured. The hyperparameter optimization ranges for different models of LightGBM, XGBoost, and CatBoost are as follows:

[0071]

[0072]

[0073]

[0074] In the above embodiments, preferably, the high-dimensional feature dataset is divided into a training set and a test set using the stratified sampling method, and the division ratio of the training set and the test set is 90% / 10%.

[0075] In the above embodiments, preferably, a preset model evaluation index is used to evaluate the trained machine learning model, and the machine learning model with the strongest generalization ability is selected as the Fenton pretreatment control membrane fouling prediction model. The specific process includes:

[0076] The mean absolute error, root mean square error, and adjusted coefficient of determination are used as the preset model evaluation indicators to evaluate each trained machine learning model;

[0077] According to the evaluation results of the mean absolute error, root mean square error, and adjusted coefficient of determination of each machine learning model, the machine learning model with the best generalization ability is selected as the Fenton pretreatment control membrane fouling prediction model;

[0078] Among them, the mean absolute error MAE, root mean square error RMSE, and adjusted coefficient of determination Adj_R 2 The calculation formulas are as follows:

[0079]

[0080] Among them, n is the number of samples in the dataset, p is the number of features, and y i is the true standard permeation flux J / J0 of the i-th sample, is the standard permeation flux J / J0 predicted by the model for the i-th sample, is the average standard permeation flux J / J0 of all samples; the lower the MAE and RMSE, and the higher the Adj_R 2 the better the prediction performance of the Fenton pretreatment control membrane fouling prediction model.

[0081] The model evaluation indicators are used to comprehensively evaluate the model prediction performance, and the model with the best generalization ability is selected as the final membrane fouling prediction model.

[0082] In the above embodiment, preferably, based on the Fenton pretreatment control membrane fouling prediction model, a genetic algorithm is used to inversely search for the optimal reagent dosage of Fenton pretreatment under the preset environmental variables. The specific process includes:

[0083] Based on the Fenton pretreatment control membrane fouling prediction model, with the standard permeation flux (J / J0) as the goal, a genetic algorithm (GA) is used to inversely search and determine the optimal Fe dosage of Fenton pretreatment for membrane fouling prevention and control under different preset environmental variables through operations such as population evolution, selection, crossover, and mutation. 2+ and H2O2 dosage.

[0084] In the above embodiment, preferably, the membrane fouling pretreatment optimization control method based on machine learning further includes:

[0085] Based on the Fenton pretreatment control membrane fouling prediction model, the test set is used to predict and verify the standard permeation flux under the condition of the optimal reagent dosage, and the optimal reagent dosage is experimentally verified in the membrane filtration system;

[0086] Perform Fenton pretreatment of membrane fouling under corresponding environmental variable conditions using the verified optimal reagent dosage to ensure efficient control of membrane fouling under different environmental conditions.

[0087] In the above embodiment, preferably, if the prediction and verification of the Fenton pretreatment control membrane fouling prediction model by the test set fails or the experimental verification fails, the candidate machine learning model is retrained, optimized, and evaluated.

[0088] The present invention also proposes an optimization control system for membrane fouling pretreatment based on machine learning, applying the method for optimizing and controlling membrane fouling pretreatment based on machine learning disclosed in any one of the above embodiments, including:

[0089] A dataset preprocessing module for constructing a high-dimensional feature dataset for preventing and controlling membrane fouling by Fenton pretreatment with different environmental variables and performing comprehensive data preprocessing on the high-dimensional feature dataset;

[0090] A candidate model training module for training and optimizing a candidate machine learning model based on a decision tree using the high-dimensional feature dataset;

[0091] An optimal model selection module for evaluating the trained machine learning model using preset model evaluation metrics and selecting the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model;

[0092] An optimal variable prediction module for inversely searching for the optimal reagent dosage of Fenton pretreatment under preset environmental variables based on the Fenton pretreatment control membrane fouling prediction model using a genetic algorithm;

[0093] An optimal variable implementation module for performing Fenton pretreatment of membrane fouling under corresponding environmental variable conditions using the optimal reagent dosage.

[0094] According to the optimization control system for membrane fouling pretreatment based on machine learning disclosed in the above embodiment, the functions to be realized by each module correspond to the steps of the method for optimizing and controlling membrane fouling pretreatment based on machine learning disclosed in the above embodiment respectively. During the implementation process, refer to the above embodiment for operation, which will not be elaborated here.

[0095] According to the method and system for optimizing and controlling membrane fouling pretreatment based on machine learning disclosed in the above embodiment, during the implementation process, the above method and system are further described in detail through the following embodiments.

[0096] Example 1

[0097] Taking the collection of research literature data related to Fenton pretreatment for controlling membrane fouling published during the period from 2010 to 2024 as an example.

[0098] A total of 2,491 experimental data on the behavior of Fenton pretreatment for preventing and controlling membrane fouling under different multivariable environments were collected from the above-mentioned relevant research literature to construct a high-dimensional feature dataset. Data related to Fenton pretreatment for controlling membrane fouling in literature pictures were all extracted using the "Engagement Digitizer" graphic digitization software. The experimental variables collected include: temperature, pressure, pH, stirring rate, flow rate, filtration time, filtration type, pollutant type, pollutant COD concentration, Fe 2+ dosage, H2O2 dosage, membrane type, membrane pore size, MWCO, initial permeation flux, and standard permeation flux J / J0 (J0 is the initial permeation flux, and J is the real-time permeation flux).

[0099] Comprehensive processing methods including data denoising, outlier removal, normalization, feature selection, and feature transformation were carried out on the constructed high-dimensional feature dataset. The data were normalized using the Z-score normalization method. For feature selection, the recursive feature elimination with cross-validation (RFEcv) and box plot correlation analysis methods were used to screen numerical and categorical features respectively. The screening results showed that pressure, stirring rate, membrane pore size, and filtration type were considered unimportant features. Except for the above 3 features, the remaining 11 features were selected as input features. Feature transformation was carried out by using the one-hot encoding method to convert categorical features into numerical features.

[0100] Construct machine learning models based on decision trees, specifically including models such as LightGBM, XGBoost, and CatBoost, which have efficient feature selection capabilities and powerful non-linear fitting capabilities. The processed data on Fenton pretreatment for controlling membrane fouling were input into different models, and the standard permeation flux J / J0 was used as the prediction target. Temperature, pH, flow rate, filtration time, pollutant type, pollutant COD concentration, Fe 2+ dosage, H2O2 dosage, membrane type, MWCO, and initial permeation flux and other 11 input features were used to train the models using the gradient boosting strategy. To prevent model overfitting, both model training and hyperparameter optimization were carried out using the five-fold cross-validation method.

[0101] First, the support vector machine (SVM) was used to investigate the influence of different splitting ratios (50% / 50%, 60% / 40%, 70% / 30%, 80% / 20%, and 90% / 10%) of the training set and the test set on the model prediction performance. The results are as Figure 2 shown. The dataset was divided using the stratified sampling method, and it was ensured that the data in the test set did not appear in the training set.Figure 2 The results show that when the training set / test set is 90% / 10%, the prediction performance of the model is the best. Therefore, the dataset is finally split with a 90% / 10% training set / test set. That is, 2,242 data are used for training and 249 data are used for testing.

[0102] During the model training process, the Optuna optimization framework based on the Bayesian Optimization algorithm is used to globally optimize the hyperparameters of each model. Through the Bayesian inference mechanism, the optimal hyperparameter combination of the model is accurately captured. The best hyperparameter combinations of different models after optimization are as follows:

[0103] The best hyperparameter combination of LightGBM is: n_estimators = 412, max_depth = 8, learning_rate = 0.02, num_leaves = 23, min_child_samples = 162, subsample = 0.87, colsample_bytree = 0.80, reg_alpha = 0.10, reg_lambda = 0.15, min_data_in_leaf = 18, bagging_fraction = 0.88, min_gain_to_split = 0.20.

[0104] The best hyperparameter combination of XGBoost is: n_estimators = 308, max_depth = 4, learning_rate = 0.012, min_child_weight = 71, gamma = 0.14, subsample: 0.63, colsample_bytree = 0.90, reg_alpha = 0.77, reg_lambda = 0.78.

[0105] The best hyperparameter combination of CatBoost is: iterations = 300, depth = 7, learning_rate = 0.07, colsample_bylevel = 0.67, bagging_temperature = 0.35, random_strength = 3.97, border_count = 249, min_child_samples = 178, leaf_estimation_method = Newton, l2_leaf_reg = 2.19.

[0106] The mean absolute error (MAE), root mean square error (RMSE) and adjusted determination coefficient (Adj_R 2)The predictive performance of the model is evaluated using three evaluation metrics, and the calculation formulas are as follows:

[0107]

[0108] where n is the number of samples in the dataset, p is the number of features, y i is the true J / J0 of the i-th sample, is the J / J0 predicted by the model for the i-th sample, is the average J / J0 of all samples. Adj_R 2 can offset the influence of the number of samples on R 2 and thus more accurately describe the goodness or badness of the model fitting. The lower the MAE and RMSE, and the higher the Adj_R 2 , the better the predictive performance of the Fenton pretreatment control membrane fouling model.

[0109] The trained Fenton pretreatment control membrane fouling model is comprehensively evaluated using the model evaluation metrics. The predictive performance evaluations of the three models under their optimal hyperparameters for Fenton pretreatment control membrane fouling are as Figure 3 shown. The results show that the CatBoost model has the best predictive performance for Fenton pretreatment control membrane fouling. Therefore, this model is finally selected as the best predictive model for Fenton pretreatment control membrane fouling.

[0110] On the basis of determining the best Fenton pretreatment control membrane fouling model, with the standard permeate flux (J / J0) as the target, the genetic algorithm (GA) is used to perform operations such as population evolution, selection, crossover, and mutation to inversely search for the optimal Fe -2 and H2O2 dosages for Fenton pretreatment during the filtration of sodium alginate (SA) through a PVDF ultrafiltration membrane at a temperature of 30°C, pH of 7, flow rate of 0, pollutant COD concentration of 34 mg / L, MWCO of 30 kDa, and initial permeate flux of 300 L·m -1 ·h 2+ . The results show that within the concentration range of 0 - 0.5 mM, the optimal Fe 2+ and H2O2 dosages for Fenton pretreatment under these conditions are 0.15 mM and 0.09 mM, respectively.

[0111] sThe Fenton pretreatment control membrane fouling model is used to predict the sodium alginate ultrafiltration membrane fouling behavior under the pretreatment conditions of 0.15 mM Fe 2+ and 0.09 mM H2O2, and this dosage is applied to the actual Fenton pretreatment sodium alginate ultrafiltration membrane filtration system for experimental system verification. The results are as Figure 4As shown, within 120 minutes of filtration, the mean square error between the experimental results and the prediction results of the CatBoost model under this condition is < 0.01%. And at this Fenton pretreatment dosage, the J / J0 of the PVDF ultrafiltration membrane for filtering SA can still reach above 0.75 after 120 minutes.

[0112] Example 2:

[0113] Based on the established optimal Fenton pretreatment membrane fouling control model, with the standard permeation flux (J / J0) as the objective, the genetic algorithm (GA) is used to perform operations such as population evolution, selection, crossover, and mutation to inversely search for the optimal Fe -2 ·h -1 dosage of Fenton pretreatment when filtering the pollutant bisphenol A (BPA) with a CK nanofiltration membrane at a temperature of 30 °C, pH of 11, flow rate of 20, pollutant COD concentration of 1000 mg / L, MWCO of 0.2 kDa, and initial permeation flux of 500 L·m 2+ The results show that within the concentration range of 0 - 0.5 mM, the optimal Fe 2+ dosage and H2O2 dosage of Fenton pretreatment under this condition are 0.08 mM and 0.049 mM respectively.

[0114] Use the Fenton pretreatment membrane fouling control model to predict the nanofiltration membrane fouling behavior of bisphenol A (BPA) under the pretreatment conditions of 0.08 mM Fe 2+ and 0.049 mM H2O2, and apply this dosage to the actual Fenton pretreatment bisphenol A (BPA) nanofiltration membrane filtration system for experimental system verification. The results are as Figure 5 shown. The mean square error between the experimental results and the prediction results of the CatBoost model under this condition is < 0.01%. And at this Fenton pretreatment dosage, the J / J0 of the CF nanofiltration membrane for filtering BPA can still reach above 0.5 after 120 minutes.

[0115] Example 3:

[0116] Based on the established optimal Fenton pretreatment membrane fouling control model, with the standard permeation flux (J / J0) as the objective, the genetic algorithm (GA) is used to perform operations such as population evolution, selection, crossover, and mutation to inversely search for the optimal Fe -2 ·h -1 dosage of Fenton pretreatment when filtering refinery wastewater (PRE) with an FP100 ultrafiltration membrane at a temperature of 25 °C, pH of 9, flow rate of 35, pollutant COD concentration of 850 mg / L, MWCO of 10 kDa, and initial permeation flux of 100 L·m 2+and H2O2 dosages. The results show that in the concentration range of 0 - 0.5 mM, the optimal Fe for Fenton pretreatment under this condition 2+ and H2O2 dosages are 0.39 mM and 0.43 mM, respectively.

[0117] Use the Fenton pretreatment to control membrane fouling model to predict the ultrafiltration membrane fouling behavior of refinery wastewater under the pretreatment conditions of 0.39 mM Fe 2+ and 0.43 mM H2O2, and apply this dosage to the actual Fenton pretreatment refinery wastewater ultrafiltration membrane filtration system for experimental system verification. The results are as Figure 6 shown. The mean square error between the experimental results and the prediction results of the CatBoost model under this condition is < 0.00%. And at this Fenton pretreatment dosage, the J / J0 of the PF100 ultrafiltration membrane for filtering PRE can still reach above 0.70 after 120 min.

[0118] Example 4:

[0119] On the basis of determining the optimal Fenton pretreatment to control membrane fouling model, with the standard permeation flux (J / J0) as the target, using the genetic algorithm (GA), through operations such as population evolution, selection, crossover, and mutation, reverse search for the optimal Fe for Fenton pretreatment when filtering papermill wastewater (PMW) with TYLG - 18 nanofiltration membrane at a temperature of 25 °C, pH of 7, flow rate of 10, pollutant COD concentration of 700 mg / L, MWCO of 5 kDa, and initial permeation flux of 1000 L·m -2 ·h -1 The results show that in the concentration range of 0 - 0.5 mM, the optimal Fe for Fenton pretreatment under this condition 2+ and H2O2 dosages are 0.47 mM and 0.15 mM, respectively. 2+ Use the Fenton pretreatment to control membrane fouling model to predict the nanofiltration membrane fouling behavior of papermill wastewater under the pretreatment conditions of 0.47 mM Fe

[0120] and 0.15 mM H2O2, and apply this dosage to the actual Fenton pretreatment papermill wastewater nanofiltration membrane filtration system for experimental system verification. The results are as 2+ shown. The mean square error between the experimental results and the prediction results of the CatBoost model under this condition is < 0.05%. And at this Fenton pretreatment dosage, the J / J0 of the TYLG - 18 nanofiltration membrane for filtering PMW can still reach above 0.35 after 120 min. Figure 7

[0121] ​The above are only the preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An optimization control method for membrane fouling pretreatment based on machine learning, characterized in that, Including: Construct a high-dimensional feature dataset for Fenton pretreatment to prevent and control membrane fouling with different environmental variables, and perform comprehensive data preprocessing on the high-dimensional feature dataset; Use the high-dimensional feature dataset to train and optimize a candidate machine learning model based on a decision tree; Evaluate the trained machine learning model using preset model evaluation metrics, and select the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model; Based on the Fenton pretreatment control membrane fouling prediction model, use a genetic algorithm to perform reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables; Use the optimal reagent dosage for Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions.

2. The method for optimizing and controlling membrane fouling pretreatment based on machine learning according to claim 1, wherein The specific process of constructing the high-dimensional feature dataset for Fenton pretreatment to prevent and control membrane fouling with different environmental variables includes: Construct a high-dimensional feature dataset of membrane fouling after Fenton pretreatment under different environmental variables, where the environmental variables include temperature, pressure, pH, flow rate, filtration time, pollutant type, pollutant concentration, COD concentration, Fe 2+ dosage, Fe 3+ dosage, H2O2 dosage, ratio of Fe to H2O2, Fenton reaction time, membrane type, MWCO, initial permeation flux, COD removal rate, and standard permeation flux J / J0; The high-dimensional feature dataset includes missing values and outliers, and the feature types include numerical and categorical features.

3. The method for optimizing and controlling the pretreatment of membrane fouling based on machine learning according to claim 2, characterized in that, The specific process of performing comprehensive data preprocessing on the high-dimensional feature dataset includes: Perform data preprocessing operations on the high-dimensional feature dataset, such as data denoising, outlier removal, data normalization, feature selection, and feature transformation.

4. The method for optimizing and controlling the pretreatment of membrane fouling based on machine learning according to claim 3, wherein The specific process of using the high-dimensional feature dataset to train and optimize a candidate machine learning model based on a decision tree includes: Select LightGBM, XGBoost, and CatBoost models based on decision trees as candidate machine learning models; Divide the high-dimensional feature dataset into a training set and a test set according to a ratio of 50% - 90% / 50% - 10%, and ensure that the data in the test set does not appear in the training set; Using the training set, five-fold cross-validation is adopted, with the standard permeate flux J / J0 as the prediction target, and temperature, pH, flow rate, filtration time, pollutant type, pollutant COD concentration, Fe 2+ dosage, H2O2 dosage, membrane type, MWCO, and initial permeate flux as 11 input features, and the candidate machine learning models are respectively trained based on the gradient boosting strategy; During the training process, use the Optuna optimization framework based on the Bayesian algorithm to globally optimize the hyperparameters of the candidate machine learning models to obtain the optimal hyperparameter combination for each model.

5. The method for optimizing and controlling membrane fouling pretreatment based on machine learning according to claim 4, wherein Use the stratified sampling method to divide the high-dimensional feature dataset into a training set and a test set, and the division ratio of the training set and the test set is 90% / 10%.

6. The method for optimizing and controlling membrane fouling pretreatment based on machine learning according to claim 5, wherein The specific process of evaluating the trained machine learning model using preset model evaluation metrics and selecting the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model includes: Use the mean absolute error, root mean square error, and corrected determination coefficient as preset model evaluation metrics to evaluate each trained machine learning model; Based on the evaluation results of the mean absolute error, root mean square error, and corrected determination coefficient of each machine learning model, select the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model; Among them, the calculation formulas of the mean absolute error MAE, the root mean square error RMSE, and the adjusted determination coefficient Adj_R 2 are as follows respectively: where n is the number of samples in the dataset, p is the number of features, and y i is the true normalized permeate flux J / J0 of the i-th sample, is the predicted normalized permeate flux J / J0 of the i-th sample by the model, is the average normalized permeate flux J / J0 of all samples; the lower the MAE and RMSE, and the higher the Adj_R 2 the better the prediction performance of the Fenton pretreatment-controlled membrane fouling prediction model.

7. The method for optimizing and controlling membrane fouling pretreatment based on machine learning according to claim 6, wherein The specific process of using a genetic algorithm to perform reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables based on the Fenton pretreatment control membrane fouling prediction model includes: Based on the Fenton pretreatment control membrane fouling prediction model, with the standard permeation flux as the target, the operation of the genetic algorithm is used to inversely search for the optimal Fe of Fenton pretreatment when preventing and controlling membrane fouling under different preset environmental variables. 2+ and H2O2 dosage.

8. The optimization control method for membrane fouling pretreatment based on machine learning according to claim 7, wherein, Also including: Based on the Fenton pretreatment control membrane fouling prediction model, use the test set to predict and verify the standard permeation flux under the condition of the optimal reagent dosage, and conduct experimental verification of the optimal reagent dosage in the membrane filtration system; Use the verified optimal reagent dosage to perform Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions.

9. The method for optimizing and controlling membrane fouling pretreatment based on machine learning according to claim 8, wherein, If the prediction and verification of the Fenton pretreatment control membrane fouling prediction model by the test set fails or the experimental verification fails, retrain, optimize, and evaluate the candidate machine learning model.

10. An optimization control system for membrane fouling pretreatment based on machine learning, characterized in that, Apply the machine learning-based membrane fouling pretreatment optimization control method according to any one of claims 1 to 9, including: A data set preprocessing module for constructing a high-dimensional feature data set for preventing and controlling membrane fouling by Fenton pretreatment with different environmental variables, and performing comprehensive data preprocessing on the high-dimensional feature data set; A candidate model training module for training and optimizing a candidate machine learning model based on a decision tree using the high-dimensional feature data set; An optimal model selection module for evaluating the trained machine learning model using a preset model evaluation index, and selecting the machine learning model with the best generalization ability as the Fenton pretreatment control membrane fouling prediction model; An optimal variable prediction module for, based on the Fenton pretreatment control membrane fouling prediction model, using a genetic algorithm to perform reverse search for the optimal reagent dosage of Fenton pretreatment under preset environmental variables; An optimal variable implementation module for using the optimal reagent dosage to perform Fenton pretreatment of membrane fouling under the corresponding environmental variable conditions.

Images