Fenton pretreatment membrane pollution prediction and key factor identification regulation and control method and system

Abstract

The invention discloses a Fenton pretreatment membrane pollution prediction and key factor identification regulation and control method and system, and the method comprises the steps: designing and carrying out a Fenton pretreatment membrane filtration coupling experiment, and collecting original experiment data; constructing a multi-dimensional feature data set of Fenton pretreatment membrane pollution behaviors; training the data set by using a deep residual network model to construct a Fenton pretreatment membrane pollution behavior prediction model; inputting the multi-dimensional features into a model for prediction; key factors influencing membrane pollution behaviors are determined by using an SHAP method; regulation and control are carried out according to key factors, optimized combination features are obtained and input into a model for feedback verification, and a closed-loop path of'experiment-prediction-explanation-optimization-verification 'is constructed. According to the technical scheme, high-precision prediction of the membrane pollution behavior is achieved, the operation efficiency of the membrane process is remarkably improved, the prediction model has good interpretability, and personalized regulation and control and directional membrane pollution relief under different water quality conditions are achieved.

Description

Technical Field

[0001] This invention relates to the field of membrane fouling technology, and in particular to a method for predicting and identifying key factors and controlling fouling in Fenton pretreatment membranes, as well as a system for predicting and identifying key factors and controlling fouling in Fenton pretreatment membranes. Background Technology

[0002] Fenton pretreatment, as a highly efficient advanced oxidation technology, has unparalleled advantages in the field of membrane fouling control, with its membrane fouling control efficiency typically being more than three times that of traditional oxidation and coagulation processes. However, the Fenton pretreatment process is highly complex, involving the synergistic effects of multiple factors, including the Fenton process itself, the physicochemical properties of fouling, membrane material characteristics, and operating conditions.

[0003] The mechanism by which Fenton pretreatment affects membrane fouling remains unclear, and related research is still largely at the macroscopic level. The optimal amount of Fenton reagent reported by different researchers to achieve the same membrane fouling control effect can vary by more than two orders of magnitude. In practical applications, achieving targeted regulation of membrane fouling through Fenton pretreatment for different pollutants and membrane materials still faces significant challenges.

[0004] Traditional single-factor experimental methods are not only limited by experimental conditions, but also require a lot of cost, time and effort, making it difficult to reveal the potential complex nonlinear relationship between experimental operating conditions, Fenton pretreatment process parameters, physicochemical properties of fouling, and membrane material properties and membrane fouling behavior. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method and system for predicting and identifying key factors influencing membrane fouling in Fenton pretreatment. By introducing a deep residual network (ResNet) model based on neural networks and training it with a multidimensional feature dataset, high-precision prediction of membrane fouling behavior after Fenton pretreatment is achieved, significantly improving the operational efficiency of the membrane process. The SHAP (Shapley Additive Explanations) method is used to quantitatively identify key factors affecting membrane fouling behavior, revealing the key factors and their mechanisms of action in the Fenton pretreatment-membrane filtration coupling process. The prediction model has good interpretability, enabling personalized control and targeted mitigation of membrane fouling under different water quality conditions.

[0006] To achieve the above objectives, this invention provides a method for predicting and identifying key factors for fouling in Fenton pretreatment membranes, comprising:

[0007] Design and implement multi-condition Fenton pretreatment membrane filtration coupling experiments and collect raw experimental data;

[0008] A multidimensional feature dataset of Fenton pretreated membrane fouling behavior was constructed based on the original experimental data.

[0009] A Fenton pretreatment membrane fouling behavior prediction model was constructed using a deep residual network model, with standard permeation flux as the target variable for model prediction, and the Fenton pretreatment membrane fouling behavior prediction model was trained using the multidimensional feature dataset.

[0010] The multidimensional features of the fouling behavior of the Fenton pretreatment membrane are used as input variables and input into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction.

[0011] The global and local SHAP values ​​are calculated using the SHAP method. The marginal contribution of the input variables to the prediction results of membrane fouling behavior is calculated from both global and local perspectives, and the key factors affecting membrane fouling behavior are identified.

[0012] By adjusting the aforementioned key factors, an optimized combination of characteristics for minimizing membrane fouling under different operating conditions is obtained;

[0013] The optimized combination features are input into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, thus constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0014] In the above technical solution, preferably, a multi-condition Fenton pretreatment membrane filtration coupling experiment is designed and implemented, and raw experimental data is collected. The specific process includes:

[0015] Design and implement Fenton pretreatment membrane filtration coupling experiments under multiple operating conditions;

[0016] Collect raw experimental data during the experiment. The raw experimental data includes one or more combinations of experimental operating conditions, Fenton pretreatment process parameters, physicochemical properties of fouling, membrane material characteristics, and membrane fouling evaluation indicators.

[0017] The experimental operating conditions include one or more combinations of filtration time, temperature, filtration pressure, membrane surface flow rate, and filtration mode.

[0018] The Fenton pretreatment process parameters include Fe 2+ Concentration, H2O2 concentration, Fe 3+ Concentration, Fe 2+ The ratio of H2O2 concentration, Fenton reaction time, stirring rate, and settling time after Fenton reaction are one or more combinations thereof;

[0019] The physicochemical properties of the pollution include one or more combinations of pollutant type, pollutant pH, total organic carbon content, pollutant chemical oxygen demand, and dissolved organic carbon.

[0020] The membrane material properties include one or more combinations of membrane type, membrane pore size, molecular weight cutoff, and initial permeation flux;

[0021] The standard permeation flux is used as the evaluation index for membrane fouling.

[0022] In the above technical solution, preferably, a multidimensional feature dataset of Fenton pretreatment membrane fouling behavior is constructed based on the original experimental data, and the specific process includes:

[0023] The collected raw experimental data are subjected to systematic preprocessing, which includes one or more combinations of missing value imputation, outlier removal, feature encoding, feature transformation, feature selection, and data normalization.

[0024] A multidimensional feature dataset of fouling behavior of Fenton pretreated membranes was constructed using experimental data after system preprocessing.

[0025] In the above technical solution, preferably, a Fenton pretreatment membrane fouling behavior prediction model is constructed using a deep residual network model, with standard permeate flux as the target variable for model prediction. The Fenton pretreatment membrane fouling behavior prediction model is trained using the multidimensional feature dataset. The specific process includes:

[0026] A stratified sampling algorithm was used to divide the multidimensional feature dataset into a training set and a test set in a 9:1 ratio;

[0027] The hyperparameters of the deep residual network model were tuned using a grid search algorithm to obtain the optimal model hyperparameter configuration.

[0028] Based on the training set, according to the optimal model hyperparameter configuration, and with standard penetration flux as the prediction target, the deep residual network model is trained to obtain the trained prediction model.

[0029] The predictive performance of the prediction model is evaluated using mean squared error and coefficient of determination based on the test set.

[0030] The prediction model is trained a preset number of times, and the prediction model with the lowest mean square error and the highest coefficient of determination is selected as the prediction model for the fouling behavior of the Fenton pretreatment membrane.

[0031] In the above technical solution, preferably, the SHAP method is used to calculate the global SHAP value and the local SHAP value, and the marginal contribution of the input variable to the membrane fouling behavior prediction result is calculated from both global and local levels to determine the key factors affecting membrane fouling behavior. The specific process includes:

[0032] The average SHAP value of all samples corresponding to each input feature is calculated as the global SHAP value. The average marginal impact of each input feature on the standard permeation flux is obtained. By ranking the importance of the input features, the top preset number of factors that have the greatest impact on membrane fouling behavior are determined as key factors.

[0033] Calculate the SHAP value of a single sample corresponding to each input feature as a local SHAP value, construct a scatter plot of SHAP value-feature value, and determine the influence relationship between the key factors and the standard penetration flux.

[0034] In the above technical solution, preferably, the key factors are controlled to obtain an optimized combination of features for minimizing membrane fouling under different operating conditions. The specific process includes:

[0035] Based on the influence relationship between the key factors and the standard permeation flux, the synergistic or inhibitory mechanisms of different key factors in the membrane fouling process are determined.

[0036] By adjusting the parameters of the key factors, and based on the optimization objective of minimizing membrane fouling, the optimal combination characteristics of Fenton pretreatment control strategies and membrane material selection under different water quality types are determined.

[0037] In the above technical solution, preferably, the optimized combination features are input into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification". The specific process includes:

[0038] The optimized combination of features is input into the Fenton pretreatment membrane fouling behavior prediction model to verify the prediction results of the optimized combination of features;

[0039] If the verification fails, the key factors are readjusted and adjusted until the verification is successful. If the verification is successful, the verification is completed, thus completing the closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0040] In the above technical solution, preferably, the missing value imputation includes median imputation for continuous variables and mode imputation for categorical variables;

[0041] The outlier removal process employs the Isolation Forest algorithm.

[0042] The feature encoding includes One-Hot encoding of categorical variables;

[0043] The feature transformation is used to construct derived variables;

[0044] The feature selection employs L1 regularization to screen features that have a significant impact on the target variable;

[0045] The data normalization uses the Z-score normalization method to normalize continuous variables.

[0046] In the above technical solution, preferably, the formula for calculating the mean square error is:

[0047]

[0048] Where MSE is the mean squared error, n is the total number of data in the test set, and i is the sequence number of the data in the test set, i = 1, 2, ..., n, y i The true standard penetration flux of the i-th data point in the test set. The predicted standard penetration flux for the i-th data point in the test set;

[0049] The formula for calculating the coefficient of determination is:

[0050]

[0051] Among them, R 2 The coefficient of determination is y = n, where n is the total number of data points in the test set, and i is the sequence number of the data points in the test set, i = 1, 2, ..., n. i The true standard penetration flux of the i-th data point in the test set. The predicted standard penetration flux for the i-th data point in the test set. This represents the mean of the test set data.

[0052] This invention also proposes a Fenton pretreatment membrane fouling prediction and key factor identification and control system, which applies the Fenton pretreatment membrane fouling prediction and key factor identification and control method disclosed in any of the above technical solutions, including:

[0053] The experimental data acquisition module is used to design and implement multi-condition Fenton pretreatment membrane filtration coupling experiments and acquire raw experimental data.

[0054] The feature data construction module is used to construct a multidimensional feature dataset of the fouling behavior of the Fenton pretreated membrane based on the original experimental data.

[0055] The prediction model training module is used to construct a Fenton pretreatment membrane fouling behavior prediction model using a deep residual network model, with standard permeation flux as the model prediction target variable, and to train the Fenton pretreatment membrane fouling behavior prediction model using the multidimensional feature dataset.

[0056] The feature variable prediction module is used to take the multidimensional features of the fouling behavior of the Fenton pretreatment membrane as input variables and input them into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction.

[0057] The key factor identification module is used to calculate the global and local SHAP values ​​using the SHAP method, and to calculate the marginal contribution of the input variables to the membrane fouling behavior prediction results from both global and local levels, thereby identifying the key factors affecting membrane fouling behavior.

[0058] The feature regulation and optimization module is used to regulate the key factors to obtain an optimized combination of features that minimize membrane fouling under different operating conditions.

[0059] The optimized feature verification module is used to input the optimized combined features into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, thus constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0060] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0061] (I) The method in this invention has the ability to predict membrane fouling behavior with high precision. By introducing a neural network model, namely a deep residual network model, and training it with a multidimensional feature dataset, it can achieve high-precision prediction of membrane fouling behavior after Fenton pretreatment. This not only significantly improves the operating efficiency of membrane processes, but also provides a replicable and scalable intelligent optimization method for membrane fouling control technology. Its prediction effect is significantly better than traditional empirical models and single-factor analysis methods.

[0062] (II) The prediction model in this invention has good interpretability; by using the SHAP method to conduct an in-depth analysis of the prediction model of Fenton pretreatment membrane fouling behavior, it is possible to quantitatively identify the key factors affecting membrane fouling behavior, reveal the key factors affecting membrane fouling behavior and their mechanisms of action in the Fenton pretreatment-membrane filtration coupling process, and provide a theoretical basis for understanding the formation mechanism of membrane fouling.

[0063] (III) The method in this invention has practical application guidance value; the method in this invention can be used to optimize membrane fouling prevention and control strategies and membrane material selection strategies, and realize personalized regulation and targeted membrane fouling mitigation under different water quality conditions.

[0064] (IV) The method in this invention helps to improve the operating efficiency and stability of membrane processes; the method in this invention can be used to assist in the construction of an intelligent membrane fouling control system, improve the stable operation of membrane separation processes, extend membrane service life, reduce operating costs, and has good engineering promotion prospects and application value. Attached Figure Description

[0065] Figure 1 This is a flowchart illustrating a method for predicting and identifying key factors for the fouling of Fenton pretreatment membranes, as disclosed in an embodiment of the present invention.

[0066] Figure 2This is a comparison chart of the actual J / J0 value and the predicted J / J0 value of the prediction model in Embodiment 1 of the present invention;

[0067] Figure 3 This is the global SHAP importance ranking diagram in Embodiment 1 of the present invention;

[0068] Figure 4 This is a local SHAP value-eigenvalue distribution map in Embodiment 1 of the present invention;

[0069] Figure 5 This is a comparison chart of the actual J / J0 value and the predicted J / J0 value of the prediction model in Embodiment 2 of the present invention;

[0070] Figure 6 This is the global SHAP importance ranking diagram in Embodiment 2 of the present invention;

[0071] Figure 7 This is a local SHAP value-feature value distribution map in Embodiment 2 of the present invention;

[0072] Figure 8 This is a comparison chart of the actual J / J0 value and the predicted J / J0 value of the prediction model in Embodiment 3 of the present invention;

[0073] Figure 9 This is the global SHAP importance ranking diagram in Embodiment 3 of the present invention;

[0074] Figure 10 This is a local SHAP value-feature value distribution diagram in Embodiment 3 of the present invention. Detailed Implementation

[0075] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] The present invention will now be described in further detail with reference to the accompanying drawings:

[0077] like Figure 1 As shown, a method for predicting and identifying key factors of fouling in a Fenton pretreatment membrane according to the present invention includes:

[0078] Design and implement multi-condition Fenton pretreatment membrane filtration coupling experiments and collect raw experimental data;

[0079] A multidimensional feature dataset of fouling behavior of Fenton pretreated membranes was constructed based on the original experimental data.

[0080] A Fenton pretreatment membrane fouling behavior prediction model was constructed using a deep residual network model. The standard permeation flux was used as the target variable for model prediction. The Fenton pretreatment membrane fouling behavior prediction model was trained using a multidimensional feature dataset.

[0081] The multidimensional features of the fouling behavior of the Fenton pretreatment membrane are used as input variables and input into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction.

[0082] The SHAP method is used to calculate global and local SHAP values, and the marginal contribution of input variables to the prediction results of membrane fouling behavior is calculated from both global and local levels to identify the key factors affecting membrane fouling behavior.

[0083] By regulating key factors, an optimized combination of characteristics for minimizing membrane fouling under different operating conditions is obtained;

[0084] The optimized combination of features is input into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, thus constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0085] In this implementation, by introducing a deep residual network model based on neural networks and training it with a multidimensional feature dataset, high-precision prediction of membrane fouling behavior after Fenton pretreatment is achieved, significantly improving the operating efficiency of the membrane process. The SHAP method is used to quantitatively identify key factors affecting membrane fouling behavior, revealing the key factors and their mechanisms of action in the Fenton pretreatment-membrane filtration coupling process. The prediction model has good interpretability, enabling personalized control and targeted membrane fouling mitigation under different water quality conditions.

[0086] Specifically, addressing the challenge that traditional single-factor experiments struggle to reveal the potentially complex nonlinear relationship between key factors and membrane fouling behavior during Fenton pretreatment's mitigation of membrane fouling, this invention innovatively introduces interpretable machine learning methods. By integrating multi-source, complex operating condition data, a model for predicting Fenton pretreatment-membrane fouling behavior and identifying key parameters is constructed. This method achieves high-precision prediction of membrane fouling behavior and identification and control of key influencing factors, effectively overcoming the limitations of traditional methods in terms of efficiency, accuracy, and adaptability. It provides theoretical support and a technical pathway for the refined application of Fenton pretreatment in membrane fouling control.

[0087] In the implementation process, a multi-condition Fenton pretreatment membrane filtration coupling experiment was first designed and implemented, collecting raw experimental data including experimental operating conditions, process parameters, and the physicochemical properties of pollutants. Based on this data, a multidimensional feature dataset containing 50 feature dimensions was constructed. A prediction model was built using a deep residual network, with a standard permeation flux (L / (m²)).2 The model is used as the prediction target. The TensorFlow framework can be used to train the model, with a training cycle of 1000 rounds and a batch size of 32. After training, the multidimensional features are input into the model for prediction. The SHAP method is used to calculate the global and local SHAP values ​​of each feature, identifying the top 5 key factors affecting membrane fouling behavior (such as Fe). 2+ Concentration, H2O2 concentration, Fe 2+ (e.g., H2O2 ratio). Based on the identification results, a multi-objective optimization algorithm is used to adjust the parameter combination of key factors to obtain an optimized combination feature scheme that minimizes membrane fouling. Finally, the optimized scheme is fed back to the prediction model to verify the effect, forming a complete "experiment-prediction-interpretation-optimization-verification" closed loop, with the entire process taking approximately 48 hours.

[0088] In the above embodiments, preferably, a multi-condition Fenton pretreatment membrane filtration coupling experiment is designed and implemented, and raw experimental data is collected. The specific process includes:

[0089] Design and implement Fenton pretreatment membrane filtration coupling experiments under multiple operating conditions;

[0090] Collect raw experimental data during the experiment. The raw experimental data includes one or more combinations of experimental operating conditions, Fenton pretreatment process parameters, physicochemical properties of fouling, membrane material characteristics, and membrane fouling evaluation indicators.

[0091] Experimental operating conditions include one or more combinations of filtration time, temperature, filtration pressure, membrane surface flow rate, and filtration mode;

[0092] Fenton pretreatment process parameters include Fe 2+ Concentration, H2O2 concentration, Fe 3+ Concentration, Fe 2+ The ratio of H2O2 concentration, Fenton reaction time, stirring rate, and settling time after Fenton reaction are one or more combinations thereof;

[0093] The physicochemical properties of pollutants include one or more combinations of pollutant type, pollutant pH, total organic carbon content, pollutant chemical oxygen demand, and dissolved organic carbon.

[0094] Membrane material properties include one or more combinations of membrane type, pore size, molecular weight cutoff, and initial permeation flux;

[0095] The standard permeation flux is used as the evaluation index for membrane fouling.

[0096] In the above embodiments, preferably, a multidimensional feature dataset of Fenton pretreatment membrane fouling behavior is constructed based on the original experimental data. The specific process includes:

[0097] The collected raw experimental data are subjected to systematic preprocessing, which includes one or more combinations of missing value imputation, outlier removal, feature encoding, feature transformation, feature selection, and data normalization.

[0098] A multidimensional feature dataset of fouling behavior of Fenton pretreated membranes was constructed using experimental data after system preprocessing.

[0099] In the above embodiments, preferably, missing value imputation includes median imputation for continuous variables and mode imputation for categorical variables;

[0100] Outlier removal is performed using the Isolation Forest algorithm.

[0101] Feature encoding includes One-Hot encoding of categorical variables;

[0102] Feature transformations are used to construct derived variables;

[0103] Feature selection employs L1 regularization to screen features that significantly influence the target variable;

[0104] Data normalization uses the Z-score standardization method to normalize continuous variables.

[0105] In the above embodiments, preferably, a Fenton pretreatment membrane fouling behavior prediction model is constructed using a deep residual network model, with standard permeate flux as the target variable for model prediction. The Fenton pretreatment membrane fouling behavior prediction model is trained using a multidimensional feature dataset. The specific process includes:

[0106] A stratified sampling algorithm was used to divide the multidimensional feature dataset into a training set and a test set in a 9:1 ratio;

[0107] The hyperparameters of the deep residual network model were tuned using a grid search algorithm to obtain the optimal model hyperparameter configuration.

[0108] Based on the training set, according to the optimal model hyperparameter configuration, and with standard penetration flux as the prediction target, the deep residual network model is trained to obtain the trained prediction model.

[0109] Based on the test set, the predictive performance of the prediction model is evaluated using mean squared error and coefficient of determination;

[0110] The prediction model was trained a predetermined number of times, and the prediction model with the lowest mean square error and the highest coefficient of determination was selected as the prediction model for the fouling behavior of the Fenton pretreatment membrane.

[0111] In the above embodiments, preferably, the formula for calculating the mean square error is:

[0112]

[0113] Where MSE is the mean squared error, n is the total number of data in the test set, and i is the sequence number of the data in the test set, i = 1, 2, ..., n, y i The true standard penetration flux of the i-th data point in the test set. The predicted standard penetration flux for the i-th data point in the test set;

[0114] The formula for calculating the coefficient of determination is:

[0115]

[0116] Among them, R 2 The coefficient of determination is y = n, where n is the total number of data points in the test set, and i is the sequence number of the data points in the test set, i = 1, 2, ..., n. i The true standard penetration flux of the i-th data point in the test set. The predicted standard penetration flux for the i-th data point in the test set. This represents the mean of the test set data.

[0117] In the above embodiments, preferably, the SHAP method is used to calculate the global SHAP value and the local SHAP value, and the marginal contribution of the input variables to the membrane fouling behavior prediction results is calculated from both global and local levels to determine the key factors affecting membrane fouling behavior. The specific process includes:

[0118] The average SHAP value of all samples corresponding to each input feature is calculated as the global SHAP value. The average marginal impact of each input feature on the standard permeation flux is obtained. By ranking the importance of the input features, the top preset number of factors that have the greatest impact on membrane fouling behavior are determined as key factors.

[0119] Calculate the SHAP value of a single sample corresponding to each input feature as the local SHAP value, construct a scatter plot of SHAP value-feature value, and determine the influence relationship between key factors and standard penetration flux.

[0120] In the above embodiments, preferably, key factors are controlled to obtain an optimized combination of characteristics for minimizing membrane fouling under different operating conditions. The specific process includes:

[0121] Based on the influence relationship between key factors and standard permeation flux, the synergistic or inhibitory mechanisms of different key factors in the membrane fouling process were determined.

[0122] By regulating the parameters of key factors, and based on the optimization objective of minimizing membrane fouling, the optimal combination characteristics of Fenton pretreatment regulation strategies and membrane material selection under different water quality types are determined.

[0123] In the above embodiments, preferably, the optimized combination of features is input into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification". The specific process includes:

[0124] The optimized combination of features was input into the Fenton pretreatment membrane fouling behavior prediction model to verify the prediction results of the optimized combination of features.

[0125] If the verification fails, the key control factors are readjusted until the verification is successful. If the verification is successful, the verification is completed, thus completing the closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0126] This invention also proposes a Fenton pretreatment membrane fouling prediction and key factor identification and control system, which applies the Fenton pretreatment membrane fouling prediction and key factor identification and control method disclosed in any of the above embodiments, including:

[0127] The experimental data acquisition module is used to design and implement multi-condition Fenton pretreatment membrane filtration coupling experiments and acquire raw experimental data.

[0128] The feature data construction module is used to construct a multidimensional feature dataset of Fenton pretreated membrane fouling behavior based on the original experimental data.

[0129] The prediction model training module is used to construct a Fenton pretreatment membrane fouling behavior prediction model using a deep residual network model. The standard permeation flux is used as the target variable for model prediction, and the Fenton pretreatment membrane fouling behavior prediction model is trained using a multidimensional feature dataset.

[0130] The feature variable prediction module is used to take the multidimensional features of the fouling behavior of the Fenton pretreatment membrane as input variables and input them into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction.

[0131] The key factor identification module is used to calculate the global and local SHAP values ​​using the SHAP method, and to calculate the marginal contribution of the input variables to the prediction results of membrane fouling behavior from both global and local levels, thereby identifying the key factors affecting membrane fouling behavior.

[0132] The feature regulation and optimization module is used to regulate key factors to obtain an optimized combination of features that minimize membrane fouling under different operating conditions.

[0133] The optimized feature verification module is used to input the optimized combined features into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, thus constructing a closed-loop path of "experiment-prediction-interpretation-optimization-verification".

[0134] The functions to be achieved by each module of the Fenton pretreatment membrane fouling prediction and key factor identification and control system disclosed in the above embodiments correspond to the steps of the Fenton pretreatment membrane fouling prediction and key factor identification and control method disclosed in the above embodiments. In the implementation process, the above embodiments are referred to for operation, and will not be repeated here.

[0135] The method and system for predicting and identifying key factors of fouling in Fenton pretreatment membranes disclosed in the above embodiments are illustrated in the following examples.

[0136] Example 1:

[0137] This embodiment presents a method for predicting fouling in Fenton pretreatment membranes and identifying and controlling key factors, such as... Figure 1 As shown, the method specifically includes the following steps:

[0138] Step 1: Collect raw experimental data:

[0139] A multi-condition Fenton pretreatment membrane filtration coupling experiment was designed and carried out. The raw experimental data collected included experimental operating conditions, Fenton pretreatment process parameters, physicochemical properties of pollutants, membrane material characteristics, and membrane fouling evaluation indicators. A total of 2491 sets of raw experimental data were collected.

[0140] The standard permeation flux is used as the evaluation index for membrane fouling. The abbreviation for standard permeation flux is J / J0. The value of standard permeation flux ranges from 0 to 1, where J is the permeation flux at time t and J0 is the initial permeation flux.

[0141] In step one, the experimental operating conditions include the filtration time (t). filt 0–240 min), temperature (T, 23–40 °C), filtration pressure (P, 1–15 bar), membrane surface velocity (Q) m 18~120L·h -1 ) and filtering modes (FM); filtering modes (FM) include dead-end filtering (DE) and cross-flow filtering (CF).

[0142] Fenton pretreatment process parameters include Fe 2+ Concentration (Fe) 2+ (0–3.6 mmol / L), H2O2 concentration (H2O2, 0–327 mmol / L), and Fe 3+ Concentration (Fe) 3+ (0–0.5 mmol / L).

[0143] The physicochemical properties of pollutants include pollutant type (FT), pollutant pH (pH, 7–11.5), and pollutant chemical oxygen demand (COD, 34–1030 mg / L); pollutant type (FT) includes bisphenol A (F... BPA ), sodium alginate (F) SA ), refining pollutants (F PRE ), reactive dyes (F AD ) and landfill leachate (F PMW ).

[0144] Membrane material properties include membrane pore size (d) pore The molecular weight cutoff (MWCO, 0.2–30 kDa) and initial permeation flux (J0, 80.2–1398.88 L·m⁻¹) are also relevant parameters. -2 ·h -1 ).

[0145] Step 2, construct a multidimensional feature dataset:

[0146] The raw experimental data collected in step one were systematically preprocessed to construct a multidimensional feature dataset of the fouling behavior of the Fenton pretreated membrane.

[0147] The system preprocessing process includes missing value imputation, outlier removal, feature encoding, feature transformation, feature selection, and data normalization.

[0148] Step two, the system preprocessing specifically includes the following procedures:

[0149] Step 201, Filling in missing values:

[0150] Median interpolation is used for continuous variables.

[0151] Step 202, Outlier Removal:

[0152] The Isolation Forest algorithm is applied, with an outlier score threshold of 0.95, to remove outliers, i.e., remove isolated data.

[0153] Step 203, Feature Encoding:

[0154] One-Hot encoding is used for categorical variable filtering mode (FM) and contaminant type (FT).

[0155] Step 204, Feature Transformation:

[0156] To construct derived variables, that is, Fe in step one 2+ The concentrations of H2O2 and Fenton are combined into a composite characteristic Fenton pretreatment concentration, which reflects the amount of Fenton reagent added.

[0157] Step 205, Feature Selection:

[0158] L1 regularization (Lasso regression) was used to screen features that significantly affected the target variable; the results retained the filtering time (t). filt ), filtration pressure (P), membrane surface velocity (Q) m ), Fenton pretreatment concentration, Fe 3+ Concentration (Fe) 3+ ), pollutant chemical oxygen demand (COD), membrane pore size (d) pore Nine features, including molecular weight cutoff (MWCO) and initial permeation flux (J0), are used as input features for the Fenton pretreatment membrane fouling behavior prediction model.

[0159] Step 206, Data Normalization:

[0160] The input features in step 205 were normalized using the Z-score normalization method (normalized data characteristics: mean 0, standard deviation 1), resulting in a multidimensional feature dataset of Fenton pretreated membrane fouling behavior suitable for modeling.

[0161] Step 3, Prediction of Fenton Pretreatment Membrane Fouling Behavior:

[0162] Using the standard permeation flux (J / J0) as the target variable for model prediction, and based on the multidimensional feature dataset of Fenton pretreatment membrane fouling behavior constructed in step two, a deep residual network model is trained to construct a Fenton pretreatment membrane fouling behavior prediction model. The Fenton pretreatment membrane fouling behavior is then predicted using the deep residual network model.

[0163] Step three involves constructing a Fenton pretreatment membrane fouling behavior prediction model, which includes the following steps:

[0164] Step 301: Divide the multidimensional feature dataset constructed in Step 2 into a training set and a test set in a 9:1 ratio, and use stratified sampling to ensure data distribution consistency.

[0165] Step 302: The hyperparameters of the deep residual network model are tuned using a grid search algorithm to obtain the optimal hyperparameter configuration of the model.

[0166] The optimal hyperparameter configuration is as follows:

[0167] Optimizer=SGD with Momentum; learning_rate=0.01; Momentum=0.92; Weight Decay=0.0001; Batch Size=256; Epochs=98; Gamma=0.12;

[0168] Step 303: Based on the training set in step 301, and according to the optimal hyperparameter configuration determined in step 302, the deep residual network model is trained with standard penetration flux as the prediction target to obtain the trained prediction model.

[0169] In steps 302 and 303, 5-fold cross-validation is used to alleviate model overfitting.

[0170] Step 304: Based on the test set from step 301, evaluate the predictive performance of the trained prediction model obtained in step 304 using mean squared error and coefficient of determination. The formula for calculating mean squared error is:

[0171]

[0172] In the formula, MSE is the mean squared error; n is the total amount of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; y i The true standard penetration flux of the i-th data point in the test set; Let be the predicted standard penetration flux for the i-th data point in the test set.

[0173] The formula for calculating the coefficient of determination is:

[0174]

[0175] In the formula, R 2 y is the coefficient of determination; n is the total number of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; i The true standard penetration flux of the i-th data point in the test set; The predicted standard penetration flux for the i-th data point in the test set; This represents the mean of the test set data.

[0176] In step 304, the coefficient of determination R 2 The higher the value, the more accurate the prediction of membrane fouling behavior; the smaller the mean square error (MSE) value, the smaller the prediction error of membrane fouling behavior.

[0177] Step 305: Repeat steps 301 to 304 five times to obtain five independent trained prediction models. Evaluate the prediction performance of these five models. Select the model with the lowest mean squared error and highest coefficient of determination on the test set as the final prediction model for Fenton pretreatment membrane fouling behavior. Figure 2 As shown, the coefficient of determination for the final prediction model is: R 2 =0.909, and the mean square error is: MSE = 0.037.

[0178] In step 305, steps 301 to 304 are repeated five times to obtain five independent trained prediction models, with the aim of reducing random errors.

[0179] Step 4, Key Factor Identification:

[0180] The interpretability analysis of the final prediction model of Fenton pretreatment membrane fouling behavior obtained in step 3 was performed using the SHAP method. The marginal contribution of each input variable to the prediction results of membrane fouling behavior was quantified from both global and local perspectives, and the key factors affecting membrane fouling behavior were identified.

[0181] In step four, the specific method for identifying key factors influencing membrane fouling behavior is as follows:

[0182] Calculate the global SHAP value to obtain the average marginal impact of each input feature on the standard permeation flux, and identify the key factors affecting membrane fouling behavior by ranking the importance of the input features; calculate the local SHAP value, construct a scatter plot of SHAP value-feature value, and identify the relationship between key factors and standard permeation flux.

[0183] Calculate the global SHAP value, and rank the SHAP features by importance as follows: Figure 3 As shown, the top 5 factors with the greatest impact on predicting membrane fouling behavior include: Fenton, Fe... 3+ P, COD and t filt Their contributions to the prediction of membrane fouling behavior were 44%, 14%, 7.5%, 7.1%, and 6.9%, respectively. Local SHAP values ​​were calculated, and further combined with SHAP value-eigenvalue distribution maps (such as...) Figure 4 As shown in the figure, the influence of various key factors on the degree of membrane fouling was analyzed, and the specific conclusions are as follows:

[0184] As the Fenton value increases, the degree of membrane fouling (the decrease in J / J0) lessens, showing a clear negative correlation trend, indicating that increasing the Fenton pretreatment concentration can effectively alleviate membrane fouling; Fe 3+ The concentration of Fe3+ is negatively correlated with the degree of membrane fouling; that is, the higher the Fe3+ concentration, the less fouling the membrane. Phosphorus (P) is positively correlated with the degree of membrane fouling; lower filtration pressure (P) helps reduce the accumulation rate of contaminants on the membrane surface, thus mitigating membrane fouling. The chemical oxygen demand (COD) of contaminants is also positively correlated with membrane fouling; that is, the lower the COD of contaminants, the less fouling the membrane. Filtration time (t)... filt It is positively correlated with membrane fouling, that is, the more the filtration time increases, the more severe the membrane fouling becomes.

[0185] This interpretability analysis enabled the quantitative identification and mechanistic elucidation of key factors in predicting the membrane fouling behavior of the Fenton pretreatment process.

[0186] Step 5, Key Factor Regulation:

[0187] Based on the key factors identified in step four, a closed-loop path of "experiment-prediction-interpretation-optimization-verification" is constructed by adjusting these key factors.

[0188] In step five, the specific method for constructing the closed-loop path of "experiment-prediction-interpretation-optimization-verification" is as follows:

[0189] Step 501: Based on the mechanism analysis results in Step 4, by regulating key factors, namely, regulating the Fenton pretreatment concentration and Fe2O3 concentration identified in Step 4... 3+ Concentration (Fe) 3+ ), filtration pressure (P), pollutant chemical oxygen demand (COD), and filtration time (t) filt These first five key factors provide optimized combinations of Fenton pretreatment control strategies and membrane material selection for different water quality types, achieving optimized operating conditions with minimal membrane fouling.

[0190] Step 502: Input the optimized combination scheme obtained in step 501 into the final prediction model obtained in step 3 for feedback verification, to ensure that the optimized combination scheme still has prediction accuracy and good pollution control effect under different water quality types, and finally construct a closed-loop process path of "experiment-prediction-interpretation-optimization-verification".

[0191] Typical operating conditions (used to illustrate the combined control effect, and do not constitute a limitation on the scope of this embodiment):

[0192] At a temperature of 25°C, the pollutant is landfill leachate (F... PWM The pollutant's chemical oxygen demand (COD) was 50 mg / L, pH was 8.5, and the membrane pore size (d) was... pore The molecular weight cutoff (MWCO) is 0.42 nm, the molecular weight cutoff (MWCO) is 20 kDa, and the initial permeation flux (J0) is 500 L·m⁻¹. -2 ·h -1 Under the operating conditions of membrane filtration:

[0193] 1. Fenton pretreatment concentration from (Fe 2+ (Fe: 1 mmol / L; H2O2: 1 mmol / L) increased to (Fe 2+ :3mmol / L; H2O2: 3mmol / L);

[0194] 2. The filtration pressure (P) was reduced from 15 bar to 5 bar;

[0195] 3. Filtering time (t) filt Reduce from 240 min to 120 min.

[0196] The optimized combination was input into the final prediction model, and the results showed that:

[0197] The standard permeate flux (J / J0) was increased from 0.62 to 0.84, and membrane fouling was reduced by 35.4%. Combined with experimental verification, the optimized measured J / J0 was finally 0.827, with a deviation of less than 5% from the model prediction, indicating that the combined control strategy has good predictive accuracy and fouling control effect under actual operating conditions.

[0198] Example 2:

[0199] This embodiment presents a method for predicting and identifying key factors and controlling fouling in Fenton pretreatment membranes, specifically targeting sodium alginate (F) under fixed operating conditions. SA ) Contaminants in polyvinylidene fluoride membranes (M PVDF Methods for predicting and identifying key factors and controlling fouling behavior of Fenton pretreatment membranes during filtration, such as... Figure 1 As shown, the method specifically includes the following steps:

[0200] Step 1: Collect raw experimental data:

[0201] We designed and conducted a multi-condition Fenton pretreatment membrane filtration coupling experiment. The raw experimental data collected included Fenton pretreatment process parameters, physicochemical properties of fouling, and membrane fouling evaluation indicators. A total of 3025 sets of raw experimental data were collected.

[0202] The standard permeation flux is used as the evaluation index for membrane fouling. The abbreviation for standard permeation flux is J / J0. The value of standard permeation flux ranges from 0 to 1, where J is the permeation flux at time t and J0 is the initial permeation flux.

[0203] In step one, the Fenton pretreatment process parameters include Fe 2+ Concentration (Fe) 2+ (0–0.2 mmol / L), H2O2 concentration (H2O2, 0–0.2 mmol / L), Fenton reaction time (t) Fenton (2.5–30 min), stirring rate (N, 100–1000 rpm), and settling time after Fenton reaction (t) stand (0-120 min).

[0204] The physicochemical properties of the pollutants include dissolved organic carbon (DOC, 5–40 mg / L).

[0205] Step 2: Perform systematic preprocessing on the raw experimental data collected in Step 1 to construct a multidimensional feature dataset of the fouling behavior of the Fenton pretreated membrane.

[0206] The system preprocessing process includes missing value imputation, outlier removal, feature transformation, feature selection, and data normalization.

[0207] Step two, the system preprocessing specifically includes the following procedures:

[0208] Step 201, Filling in missing values:

[0209] Median interpolation is used for continuous variables.

[0210] Step 202, Outlier Removal:

[0211] The Isolation Forest algorithm is applied, with an outlier score threshold of 0.95, to remove outliers, i.e., remove isolated data.

[0212] Step 203, Feature Transformation:

[0213] The derivation variable is constructed by converting the dissolved organic carbon (DOC) in step one into the ratio of composite characteristic dissolved organic carbon to Fenton pretreatment concentration (DOC / Fenton), where the Fenton pretreatment concentration is derived from the Fe in step one. 2+ It is a combination of the concentration and the H2O2 concentration.

[0214] Step 204, Feature Selection:

[0215] L1 regularization (Lasso regression) was used to screen features that significantly affected the target variable; the results showed 6 features (Fe 2+ H2O2, t Fenton N, t stand Both DOC / Fenton were retained as input features for the Fenton pretreatment membrane fouling behavior prediction model.

[0216] Step 205, Data Normalization:

[0217] The input features in step 204 were normalized using the Z-score normalization method (normalized data characteristics: mean 0, standard deviation 1), resulting in a multidimensional feature dataset of Fenton pretreated membrane fouling behavior suitable for modeling.

[0218] Step 3, Prediction of Fenton Pretreatment Membrane Fouling Behavior:

[0219] Using standard permeation flux as the target variable for model prediction, and based on the multidimensional feature dataset of Fenton pretreatment membrane fouling behavior constructed in step two, a deep residual network model is trained to construct a Fenton pretreatment membrane fouling behavior prediction model. The Fenton pretreatment membrane fouling behavior is then predicted using the deep residual network model.

[0220] Step three involves constructing a Fenton pretreatment membrane fouling behavior prediction model, which includes the following steps:

[0221] Step 301: Divide the multidimensional feature dataset constructed in Step 2 into a training set and a test set in a 9:1 ratio, and use stratified sampling to ensure data distribution consistency.

[0222] Step 302: The hyperparameters of the base-depth residual network model are tuned using a grid search algorithm to obtain the optimal hyperparameter configuration of the model.

[0223] The optimal hyperparameter configuration is as follows:

[0224] Optimizer=AdamW; learning_rate=0.004; Momentum=0.82; Weight Decay=0.0004; Batch Size=128; Epochs=64; Gamma=0.25.

[0225] Step 303: Based on the training set in step 301, and according to the optimal hyperparameter configuration determined in step 302, the deep residual network model is trained with standard penetration flux as the prediction target to obtain the trained prediction model.

[0226] In steps 302 and 303, 5-fold cross-validation is used to alleviate model overfitting.

[0227] Step 304: Based on the test set from step 301, evaluate the predictive performance of the trained prediction model obtained in step 303 using mean squared error and coefficient of determination. The formula for calculating mean squared error is:

[0228]

[0229] In the formula, MSE is the mean squared error; n is the total amount of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; y i The true standard penetration flux of the i-th data point in the test set; Let be the predicted standard penetration flux for the i-th data point in the test set.

[0230] The formula for calculating the coefficient of determination is:

[0231]

[0232] In the formula, R 2 y is the coefficient of determination; n is the total number of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; i The true standard penetration flux of the i-th data point in the test set; The predicted standard penetration flux for the i-th data point in the test set; This represents the mean of the test set data.

[0233] In step 304, the coefficient of determination R 2 The higher the value, the more accurate the prediction of membrane fouling behavior; the smaller the mean square error (MSE) value, the smaller the prediction error of membrane fouling behavior.

[0234] Step 305: Repeat steps 301 to 304 five times to obtain five independent trained prediction models, and evaluate the prediction performance of the five independent trained prediction models; select the trained prediction model with the lowest mean squared error and the highest coefficient of determination on the test set as the final prediction model for the fouling behavior of the Fenton pretreatment membrane; such as Figure 5 As shown, the coefficient of determination for the final prediction model is: R 2 =0.971; the lowest mean square error is: MSE = 0.022.

[0235] In step 305, steps 301 to 304 are repeated five times to obtain five trained prediction models, with the aim of reducing random errors.

[0236] Step 4, Key Factor Identification:

[0237] The interpretability analysis of the final prediction model of Fenton pretreatment membrane fouling behavior obtained in step 3 was performed using the SHAP method. The marginal contribution of each input variable to the prediction results of membrane fouling behavior was quantified from both global and local perspectives, and the key factors affecting membrane fouling behavior were identified.

[0238] In step four, the specific method for identifying key factors influencing membrane fouling behavior is as follows:

[0239] Calculate the global SHAP value to obtain the average marginal impact of each input feature on the standard permeation flux, and identify the key factors affecting membrane fouling behavior by ranking the importance of the input features; calculate the local SHAP value, construct a scatter plot of SHAP value-feature value, and identify the relationship between key factors and standard permeation flux.

[0240] Calculate the global SHAP value, and rank the SHAP features by importance as follows: Figure 6 As shown, the top 5 factors with the greatest impact on predicting membrane fouling behavior include: Fe 2+ Concentration (Fe) 2+ The ratio of dissolved organic carbon to Fenton pretreatment concentration (DOC / Fenton), H2O2 concentration (H2O2), and Fenton reaction time (t) are also considered. Fenton ) and the settling time after the Fenton reaction (t) stand The contributions of these factors to the prediction of membrane fouling behavior were 34.5%, 23.5%, 19%, 8.8%, and 8.7%, respectively. Local SHAP values ​​were calculated, and further analysis was performed using SHAP value-eigenvalue distribution maps (such as...). Figure 7As shown in the figure, the influence of various key factors on the degree of membrane fouling was analyzed, and the specific conclusions are as follows:

[0241] Fe 2+ As the Fe value increases, the degree of membrane fouling (the decrease in J / J0) decreases, showing a clear negative correlation trend, indicating that increasing Fe in the Fenton pretreatment process reduces the degree of membrane fouling. 2+ Concentration can effectively alleviate membrane fouling; DOC / Fenton ratio is positively correlated with the degree of membrane fouling, the smaller the DOC / Fenton value, the less membrane fouling, that is, at a unit Fenton pretreatment concentration, the less dissolved organic carbon, and the better the Fenton pretreatment effect on mitigating membrane fouling; H2O2 also shows a negative correlation with the degree of membrane fouling, the higher the H2O2 concentration, the better the membrane fouling mitigation effect; Fenton reaction time (t Fenton ), standing time after Fenton reaction (t) stand Both the stirring rate (N) during the Fenton reaction and the degree of membrane fouling show a negative correlation, with a higher value being more beneficial to mitigating membrane fouling.

[0242] This interpretability analysis enabled the quantitative identification and mechanistic elucidation of key factors in predicting the membrane fouling behavior of the Fenton pretreatment process.

[0243] Step 5, Key Factor Regulation:

[0244] Based on the key factors identified in step four, a closed-loop path of "experiment-prediction-interpretation-optimization-verification" is constructed by adjusting these key factors.

[0245] In step five, the specific method for constructing the closed-loop path of "experiment-prediction-interpretation-optimization-verification" is as follows:

[0246] Step 501: Based on the mechanism analysis results in Step 4, the key factors are controlled, namely, the Fe identified in Step 4. 2+ Concentration (Fe) 2+ The ratio of dissolved organic carbon to Fenton pretreatment concentration (DOC / Fenton), H2O2 concentration (H2O2), and Fenton reaction time (t) are also considered. Fenton ) and the settling time after the Fenton reaction (t) stand Based on these five key factors, we propose optimized combinations for different operating conditions to achieve optimal operating conditions that minimize membrane fouling.

[0247] Step 502: Input the optimized combination scheme obtained in step 501 into the final prediction model of Fenton pretreatment membrane fouling behavior obtained in step 3 for feedback verification, to ensure that the optimized combination scheme still has prediction accuracy and good fouling control effect under different operating conditions, and finally construct a closed-loop process path of "experiment-prediction-interpretation-optimization-verification".

[0248] Typical operating conditions (used to illustrate the combined control effect, and do not constitute a limitation on the scope of this embodiment):

[0249] The pollutant is sodium alginate (F) SA The dissolved organic carbon (DOC) concentration was 10 mg / L, the Fenton reaction stirring rate (N) was 500 rpm, and the membrane material was polyvinylidene fluoride membrane (M). PVDF Under the operating conditions of )

[0250] 1. Fe 2+ The concentration was increased from 0.05 mmol / L to 0.2 mmol / L;

[0251] 2. The H2O2 concentration increased from 0.05 mmol / L to 0.2 mmol / L;

[0252] 3. The ratio of dissolved organic carbon to Fenton pretreatment concentration (DOC / Fenton) decreased from 200 to 50;

[0253] 4. Fenton reaction time (t) Fenton Extend the time from 5 minutes to 20 minutes;

[0254] 5. Settling time after Fenton reaction (t) stand Extend the time from 0 min to 120 min;

[0255] The optimized combination was input into the final prediction model, and the results showed that:

[0256] The standard permeate flux (J / J0) was increased from 0.59 to 0.88, and membrane fouling was reduced by 49.1%. Combined with experimental verification, the optimized measured J / J0 was finally 0.893, with a deviation of less than 5% from the model prediction, indicating that the combined control strategy has good predictive accuracy and fouling control effect under actual operating conditions.

[0257] Example 3:

[0258] This embodiment presents a method for predicting fouling in Fenton pretreatment membranes and identifying and controlling key factors, such as... Figure 1 As shown, the method specifically includes the following steps:

[0259] Step 1: Collect raw experimental data:

[0260] We designed and conducted a multi-condition Fenton pretreatment membrane filtration coupling experiment, collecting a total of 6,534 sets of raw experimental data, including experimental operating conditions, Fenton pretreatment process parameters, physicochemical properties of pollutants, membrane material characteristics, and membrane fouling evaluation indicators.

[0261] The standard permeation flux is used as the evaluation index for membrane fouling. The abbreviation for standard permeation flux is J / J0. The value of standard permeation flux ranges from 0 to 1, where J is the permeation flux at time t and J0 is the initial permeation flux.

[0262] In step one, the experimental operating conditions include the filtration time (t). filt , 0–120 min), temperature (T, 10–30 °C) and filtration pressure (P, 3–10 bar).

[0263] Fenton pretreatment process parameters include Fe 2+ Concentration (Fe) 2+ (0–0.3 mmol / L), H2O2 concentration (H2O2, 0–0.2 mmol / L), Fe 2+ The ratio of Fe to H2O2 concentration 2+ / H2O2, 0~2), Fenton reaction time (t) Fenton (2.5–40 min), stirring rate (N, 0–500 rpm), and settling time after Fenton reaction (t) stand (0–160 min).

[0264] The physicochemical properties of pollutants include pollutant type (FT) and total organic carbon content (TOC, 10–40 mg / L); pollutant type (FT) includes bovine serum albumin (F... BSA ), sodium alginate (F) SA ) and humic acid (F HA ).

[0265] Membrane material properties include membrane type (MT) and molecular weight cutoff (MWCO, 30–500 kDa); membrane type (MT) includes polyvinylidene fluoride membrane (M... PVDF ) and polyethersulfone membrane (M PES ).

[0266] Step 2, construct a multidimensional feature dataset:

[0267] The raw experimental data collected in step one were systematically preprocessed to construct a multidimensional feature dataset of the fouling behavior of the Fenton pretreated membrane.

[0268] The system preprocessing process includes missing value imputation, outlier removal, feature encoding, feature selection, and data normalization.

[0269] Step two, the system preprocessing specifically includes the following procedures:

[0270] Step 201, Filling in missing values:

[0271] Median interpolation is used for continuous variables.

[0272] Step 202, Outlier Removal:

[0273] The Isolation Forest algorithm is applied, with an outlier score threshold of 0.95, to remove outliers, i.e., remove isolated data.

[0274] Step 203, Feature Encoding:

[0275] One-Hot coding was used for the categorical variables pollutant type (FT) and membrane type (MT).

[0276] Step 204, Feature Selection:

[0277] L1 regularization (Lasso regression) was used to screen features that significantly affected the target variable; the results retained the filtering time (t). filt ), Fe 2+ Concentration (Fe) 2+ H2O2 concentration (H2O2), Fe 2+ The ratio of Fe to H2O2 concentration 2+ / H2O2), Fenton reaction time (t) Fenton ), stirring rate (N), and settling time after Fenton reaction (t) stand Total organic carbon (TOC), molecular weight cutoff (MWCO), bovine serum albumin (F) BSA ), sodium alginate (F) SA ), humic acid (F) HA ), polyvinylidene fluoride membrane (M PVDF ) and polyethersulfone membrane (M PES These 14 features serve as input features for the Fenton pretreatment membrane fouling behavior prediction model.

[0278] Step 205, Data Normalization:

[0279] The input features in step 204 were normalized using the Z-score normalization method (normalized data characteristics: mean 0, standard deviation 1), resulting in a multidimensional feature dataset of Fenton pretreated membrane fouling behavior suitable for modeling.

[0280] Step 3, Prediction of Fenton Pretreatment Membrane Fouling Behavior:

[0281] Using standard permeation flux as the target variable for model prediction, and based on the multidimensional feature dataset of Fenton pretreatment membrane fouling behavior constructed in step two, a deep residual network model is trained to construct a Fenton pretreatment membrane fouling behavior prediction model. The Fenton pretreatment membrane fouling behavior is then predicted using the deep residual network model.

[0282] Step three involves constructing a Fenton pretreatment membrane fouling behavior prediction model, which includes the following steps:

[0283] Step 301: Divide the multidimensional feature dataset constructed in Step 2 into a training set and a test set in a 9:1 ratio, and use stratified sampling to ensure data distribution consistency.

[0284] Step 302: The hyperparameters of the deep residual network model are tuned using a grid search algorithm to obtain the optimal hyperparameter configuration of the model.

[0285] The optimal hyperparameter configuration is as follows:

[0286] Optimizer=SGD with Momentum; learning_rate=0.017; Momentum=0.085; Weight Decay=0.0025; Batch Size=256; Epochs=69; Gamma=0.46.

[0287] Step 303: Based on the training set in step 301, and according to the optimal hyperparameter configuration determined in step 302, the deep residual network model is trained with standard penetration flux as the prediction target to obtain the trained prediction model.

[0288] In steps 302 and 303, 5-fold cross-validation is used to alleviate model overfitting.

[0289] Step 304: Based on the test set from step 301, evaluate the predictive performance of the trained prediction model obtained in step 303 using mean squared error and coefficient of determination. The formula for calculating mean squared error is:

[0290]

[0291] In the formula, MSE is the mean squared error; n is the total amount of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; y i The true standard penetration flux of the i-th data point in the test set; Let be the predicted standard penetration flux for the i-th data point in the test set.

[0292] The formula for calculating the coefficient of determination is:

[0293]

[0294] In the formula, R 2 y is the coefficient of determination; n is the total number of data in the test set; i is the sequence number of the data in the test set, i = 1, 2, ..., n; i The true standard penetration flux of the i-th data point in the test set; The predicted standard penetration flux for the i-th data point in the test set; This represents the mean of the test set data.

[0295] In step 304, the coefficient of determination R 2 The higher the value, the more accurate the prediction of membrane fouling behavior; the smaller the mean square error (MSE) value, the smaller the prediction error of membrane fouling behavior.

[0296] Step 305: Repeat steps 301 to 304 five times to obtain five independent trained prediction models, and evaluate the prediction performance of the five independent trained prediction models; select the trained prediction model with the lowest mean squared error and the highest coefficient of determination on the test set as the final prediction model for the fouling behavior of the Fenton pretreatment membrane; such as Figure 8 As shown, the coefficient of determination for the final prediction model is: R 2 =0.877, and the mean square error is: MSE = 0.045.

[0297] In step 305, steps 301 to 304 are repeated five times to obtain five trained prediction models, with the aim of reducing random errors.

[0298] Step 4, Key Factor Identification:

[0299] The interpretability analysis of the final prediction model of Fenton pretreatment membrane fouling behavior obtained in step 3 was performed using the SHAP method. The marginal contribution of each input variable to the prediction results of membrane fouling behavior was quantified from both global and local perspectives, and the key factors affecting membrane fouling behavior were identified.

[0300] In step four, the specific method for identifying key factors influencing membrane fouling behavior is as follows:

[0301] Calculate the global SHAP value to obtain the average marginal impact of each input feature on the standard permeation flux, and identify the key factors affecting membrane fouling behavior by ranking the importance of the input features; calculate the local SHAP value, construct a scatter plot of SHAP value-feature value, and identify the relationship between key factors and standard permeation flux.

[0302] Calculate the global SHAP value, and rank the SHAP features by importance as follows: Figure 9 As shown, the top 5 factors with the greatest impact on predicting membrane fouling behavior include: Fe 2+ t filt TOC, H2O2 and F BSA Their contributions to the prediction of membrane fouling behavior were 23.1%, 15.4%, 12.7%, 11.6%, and 8.3%, respectively. Local SHAP values ​​were calculated, and further combined with SHAP value-eigenvalue distribution maps (such as...) Figure 10 As shown in the figure, the influence of various key factors on the degree of membrane fouling was analyzed, and the specific conclusions are as follows:

[0303] With Fe2+ As the value increases, the degree of membrane fouling (the decrease in J / J0) decreases, showing a clear negative correlation trend, indicating that increasing Fe in the Fenton pretreatment process reduces the degree of membrane fouling. 2+ Concentration can effectively alleviate membrane fouling; filtration time (t) filt ) is positively correlated with membrane fouling, t filt The lower the value, the less membrane fouling; total organic carbon (TOC) content is also positively correlated with membrane fouling, with lower TOC levels indicating less membrane fouling; H₂O₂ is negatively correlated with the degree of membrane fouling, with higher H₂O₂ concentrations resulting in better mitigation of membrane fouling; F BSA It is positively correlated with membrane fouling, F BSA The more membranes there are, the more severe the fouling becomes.

[0304] This interpretability analysis enabled the quantitative identification and mechanistic elucidation of key factors in predicting the membrane fouling behavior of the Fenton pretreatment process.

[0305] Step 5, Key Factor Regulation:

[0306] Based on the key factors identified in step four, a closed-loop path of "experiment-prediction-interpretation-optimization-verification" is constructed by adjusting these key factors.

[0307] In step five, the specific method for constructing the closed-loop path of "experiment-prediction-interpretation-optimization-verification" is as follows:

[0308] Step 501: Based on the mechanism analysis results in Step 4, the key factors are controlled, namely, the Fe identified in Step 4. 2+ Concentration (Fe) 2+ ), Filtering time (t) filt Total organic carbon (TOC), H2O2 concentration (H2O2), and bovine serum albumin (F) BSA Based on these five key factors, we propose optimized combinations of Fenton pretreatment control strategies and membrane material selection under different operating conditions to achieve optimized operating conditions with minimal membrane fouling.

[0309] Step 502: Input the optimized combination scheme obtained in step 501 into the final prediction model of Fenton pretreatment membrane fouling behavior obtained in step 3 for feedback verification, to ensure that the optimized combination scheme still has prediction accuracy and good fouling control effect under different operating conditions, and finally construct a closed-loop process path of "experiment-prediction-interpretation-optimization-verification".

[0310] Typical operating conditions (used to illustrate the combined control effect, and do not constitute a limitation on the scope of this embodiment):

[0311] At a temperature (T) of 30℃, the pollutant is humic acid (F HAThe filtration pressure (P) is 5 bar, and the membrane type is M. PVDF The molecular weight cutoff (MWCO) is 100 kDa, and the Fenton reaction time (t) is... Fenton The reaction time was 10 min, the stirring rate (N) was 300 rpm, and the settling time after the Fenton reaction was t. stand ) for 50 min, Fe 2+ The ratio of Fe to H2O2 concentration 2+ Under membrane filtration operating conditions where H2O2 is 1.5:

[0312] 1. Fe 2+ The concentration was increased from 0.15 mmol / L to 0.3 mmol / L;

[0313] 2. The H2O2 concentration was increased from 0.1 mmol / L to 0.2 mmol / L;

[0314] 3. The total organic carbon content of pollutants decreased from 40 mg / L to 10 mg / L;

[0315] 4. Filtering time (t) filt Reduce the time from 120 min to 40 min.

[0316] The optimized combination was input into the final prediction model, and the results showed that:

[0317] The standard permeate flux (J / J0) increased from 0.71 to 0.89, and membrane fouling decreased by 25.3%. Combined with experimental verification, the optimized measured J / J0 was finally 0.902, with a deviation of less than 5% from the model prediction, indicating that the combined control strategy has good predictive accuracy and fouling control effect under actual operating conditions.

[0318] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A Fenton pretreatment membrane fouling prediction and key factor identification regulation method, characterized in that, The application relates to a method for predicting and optimizing membrane fouling behavior in Fenton pretreatment membrane filtration coupling experiments. The method comprises the following steps: designing and implementing Fenton pretreatment membrane filtration coupling experiments under multiple working conditions, and collecting original experimental data; constructing a multi-dimensional feature data set of Fenton pretreatment membrane fouling behavior based on the original experimental data; constructing a Fenton pretreatment membrane fouling behavior prediction model by using a deep residual network model, taking standard permeation flux as a model prediction target variable, and training the Fenton pretreatment membrane fouling behavior prediction model by using the multi-dimensional feature data set; inputting multi-dimensional features of Fenton pretreatment membrane fouling behavior as input variables into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction; calculating global SHAP values and local SHAP values by using a SHAP method, calculating marginal contributions of the input variables to the prediction results of membrane fouling behavior from global and local two aspects, and determining key factors affecting membrane fouling behavior; controlling the key factors to obtain optimized combination features of membrane fouling minimization under different working conditions; 2. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 1, characterized in that, inputting the optimized combination features into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, and constructing a closed-loop path of "experiment-prediction-explanation-optimization-verification". The method comprises the following steps: designing and implementing Fenton pretreatment membrane filtration coupling experiments under multiple working conditions, and collecting original experimental data, and the specific process comprises the following steps: designing Fenton pretreatment membrane filtration coupling experiments under multiple working conditions and implementing the experiments; the Fenton pre-treatment process parameters include Fe 2+ concentration, H2O2 concentration, Fe 3+ concentration, Fe 2+ one or more combinations of the ratio of H2O2 concentration to Fe concentration, Fenton reaction time, stirring rate, and post-Fenton reaction standing time; collecting original experimental data in the experimental process, wherein the original experimental data comprise one or more combinations of experimental operation conditions, Fenton pretreatment process parameters, pollution physical and chemical properties, membrane material characteristics and membrane fouling evaluation indexes; the experimental operation conditions comprise one or more combinations of filtration time, temperature, filtration pressure, membrane surface flow rate and filtration mode; the pollution physical and chemical properties comprise one or more combinations of pollutant type, pollutant pH, total organic carbon content, pollutant chemical oxygen demand and dissolved organic carbon; 3. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 2, characterized in that, the membrane material characteristics comprise one or more combinations of membrane type, membrane pore size, molecular weight cut-off and initial permeation flux; the membrane fouling evaluation index adopts standard permeation flux. The method comprises the following steps:

4. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 3, characterized in that, performing system preprocessing on the collected original experimental data, wherein the system preprocessing comprises one or more combinations of missing value filling, abnormal value elimination, feature encoding, feature conversion, feature selection and data normalization; constructing a multi-dimensional feature data set of Fenton pretreatment membrane fouling behavior by using the system-preprocessed experimental data. The method comprises the following steps: adopting a stratified sampling algorithm to divide the multi-dimensional feature data set into a training set and a test set according to a 9:1 ratio; adopting a grid search algorithm to optimize hyperparameters of the deep residual network model, and obtaining optimal model hyperparameter configurations; Based on the training set, the deep residual network model is trained according to the optimal model hyperparameter configuration, with the standard permeation flux as the prediction target, to obtain a trained prediction model; According to the test set, the prediction performance of the prediction model is evaluated by using the mean square error and the determination coefficient; The prediction model is trained for a preset number of times, and the prediction model with the lowest mean square error and the highest determination coefficient is selected as the Fenton pretreatment membrane pollution behavior prediction model.

5. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 4, characterized in that, The SHAP method is used to calculate the global SHAP value and the local SHAP value, and the marginal contribution of the input variables to the prediction result of the membrane pollution behavior is calculated from the global and local two levels to determine the key factors affecting the membrane pollution behavior, and the specific process includes: The average SHAP value of all samples corresponding to each input feature is calculated as the global SHAP value, the average marginal influence of each input feature on the standard permeation flux is obtained, and the top preset number of factors that affect the membrane pollution behavior most are determined as the key factors by sorting the importance of the input features; The SHAP value of a single sample corresponding to each input feature is calculated as the local SHAP value, a SHAP value-feature value scatter plot is constructed, and the influence relationship between the key factors and the standard permeation flux is determined.

6. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 5, characterized in that, The key factors are regulated to obtain the optimized combination features of the membrane pollution minimization optimization objective under different working conditions, and the specific process includes: Based on the influence relationship between the key factors and the standard permeation flux, the synergistic or inhibitory mechanism of different key factors in the membrane pollution formation process is determined; The parameters of the key factors are regulated, and the optimized combination features of the Fenton pretreatment regulation strategy and the membrane material selection under different water quality types are determined based on the optimization objective of membrane pollution minimization.

7. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 6, characterized in that, The optimized combination features are input into the Fenton pretreatment membrane pollution behavior prediction model for feedback verification, and a closed-loop path of "experiment-prediction-explanation-optimization-verification" is constructed, and the specific process includes: The optimized combination features are input into the Fenton pretreatment membrane pollution behavior prediction model, and the prediction result of the optimized combination features is verified; If the verification fails, the key factors are adjusted and regulated again until the verification passes, and if the verification passes, the verification is completed, and the closed-loop path of "experiment-prediction-explanation-optimization-verification" is completed.

8. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 3, characterized in that, The missing value filling includes using the median for continuous variable interpolation and using the mode for category variable interpolation; The outlier rejection applies the Isolation Forest algorithm for outlier rejection; The feature encoding includes One-Hot encoding for category variables; The feature transformation is used for derivative variable construction; The feature selection uses the L1 regularization method to screen the features that have a significant impact on the target variable; The data normalization uses the Z-score standardization method to normalize the continuous variables.

9. The Fenton pre-treatment membrane fouling prediction and key factor identification regulation method according to claim 4, characterized in that, The calculation formula of the mean square error is: wherein MSE is mean square error, n is the total amount of data of the test set, i is the serial number of the data of the test set, i = 1, 2, …, n, y i is the true standard permeation flux of the i th data of the test set, is the predicted standard permeation flux of the i th data of the test set; The calculation formula of the determination coefficient is: wherein R 2 is the determination coefficient, n is the total amount of data of the test set, i is the sequence number of the data of the test set, i = 1, 2, …, n, y i is the true standard permeation flux of the i th data of the test set, is the predicted standard permeation flux of the i th data of the test set, is the mean value of the data of the test set.

10. A Fenton pretreatment membrane fouling prediction and key factor identification regulation system, characterized in that, The Fenton pretreatment membrane pollution prediction and key factor identification regulation method according to any one of claims 1-9, comprising: An experimental data collection module is configured to design and implement Fenton pretreatment membrane filtration coupling experiments under multiple working conditions and collect original experimental data; A feature data construction module is configured to construct a multi-dimensional feature data set of Fenton pretreatment membrane fouling behavior based on the original experimental data; A prediction model training module is configured to construct a Fenton pretreatment membrane fouling behavior prediction model using a deep residual network model, take standard permeate flux as a model prediction target variable, and train the Fenton pretreatment membrane fouling behavior prediction model using the multi-dimensional feature data set; A feature variable prediction module is configured to input the multi-dimensional features of Fenton pretreatment membrane fouling behavior as input variables into the trained Fenton pretreatment membrane fouling behavior prediction model for prediction; A key factor identification module is configured to calculate global SHAP values and local SHAP values using the SHAP method, calculate the marginal contribution of the input variables to the prediction results of membrane fouling behavior from the global and local levels, and determine the key factors affecting membrane fouling behavior; A feature regulation and optimization module is configured to regulate the key factors to obtain an optimized combination feature of the membrane fouling minimization optimization objective under different working conditions; An optimized feature verification module is configured to input the optimized combination feature into the Fenton pretreatment membrane fouling behavior prediction model for feedback verification, and construct a closed-loop path of "experiment-prediction-explanation-optimization-verification".

Images