Intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback

Abstract

The present invention discloses an intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback, constructs a multi-stage prediction architecture adapted to the characteristics of source water, corrects subsequent prediction values ​​through feedback of the preceding reagent dosage, and suppresses error transmission; at the same time, based on the intermittent dosing characteristics of PAC, a dynamic classification regression mechanism of classification judgment + regression prediction is established; an efficient data governance method for multi-system and multi-precision data of a water plant is designed to solve the outlier processing and feature engineering optimization in a mixed scenario of minute-level and daily precision data; a lightweight model based on XGBoost is developed, which significantly improves the adaptability to source water and the robustness of the model, and provides a high-precision, explainable, and easy-to-deploy intelligent precision dosing solution for source water plants, promoting the transformation of the water treatment industry towards intelligent, efficient, low-consumption, and low-carbon development.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control of water treatment processes, and in particular to an intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback mechanism. Background Art

[0002] In recent years, with the development of industry and agriculture, the large-scale discharge of chemical wastewater, domestic sewage, etc. and the large-scale use of fertilizers, pesticides, herbicides, etc. have made the pollution and eutrophication of surface water increasingly serious, and the seasonal changes in water quality and quantity are significant. Traditional coagulation, sedimentation, filtration, and disinfection processes have poor control effects on ammonia nitrogen, trace organic matter, dissolved organic matter, odor, etc., and the treated water is difficult to meet the relevant national standards for drinking water quality. Therefore, removing ammonia nitrogen, organic micropollutants, dissolved organic matter, odor, etc. in water has become a difficult problem that water supply plants urgently need to solve.

[0003] Coagulation and sedimentation is a commonly used water purification technology. This technique uses chemicals to destabilize colloids and fine suspended matter in the treated water, coagulating them into flocs. The flocs are then separated to purify the water. Coagulation and dosing are a core step in the water purification process, directly impacting effluent quality and operating costs. Traditional water plants rely on manual experience combined with beaker tests to adjust dosing. This can lead to problems such as poor real-time performance, delayed control, and overdosing, making it difficult to adapt to dynamically changing water quality and process conditions.

[0004] In recent years, although machine learning technology has shown potential in coagulation dosing, it still has certain application limitations: (1) The lack of multi-agent synergistic prediction. Existing studies mostly build single-agent dosage prediction models based on water quality parameters, ignoring the temporal correlation of multi-agent synergistic dosing (such as pre-chlorination and dual-agent combination), and cannot achieve accurate control of the entire chain of "raw water-dosing-effluent". In the intermittent dosing scenario of stopping PAC during the low temperature and low turbidity period in winter, the distribution of PAC dosage data is discontinuous, and directly using traditional regression model modeling and prediction will lead to a surge in errors. (2) The bottleneck of low-quality data modeling. The actual data of water plants has problems such as mixed minute-level and daily accuracy, abnormal or missing sensor data, etc., resulting in poor model robustness and insufficient dynamic adaptability.

[0005] Therefore, there is an urgent need to develop an intelligent drug delivery method that integrates multi-target prediction, dynamic classification regression and low-quality data robustness optimization. Summary of the Invention

[0006] Purpose of the invention: In view of the defects of the existing technology, the purpose of the present invention is to provide an intelligent low-carbon coagulant dosing method based on multi-stage machine learning and dynamic feedback, which can realize the accurate prediction and optimization of the coagulant dosage of the water plant through a multi-stage modeling framework, a classification-regression coupling mechanism and dynamic data governance technology.

[0007] Technical solution:

[0008] On the one hand, the present invention provides an intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback, characterized in that it includes the following steps:

[0009] S1: Multi-source data acquisition and fusion: Obtain real-time raw water quality parameters, flow rate, dosage data, daily water quality data and effluent turbidity from the water plant SCADA system and integrated information management system;

[0010] S2: Dynamic data management: Hourly data is converted to daily data, the median is resampled, and the data is merged with the daily inspection data. Missing values ​​and outliers are processed.

[0011] S3: Construction of multi-stage prediction framework: Establishing the chain prediction logic of pre-chlorination NaClO→FeCl3→PAC→effluent turbidity, with the dosage of the preceding reagents as the input feature of the subsequent model;

[0012] S4: Dynamic feedback optimization: Based on the effluent turbidity prediction results, if the predicted value exceeds the threshold, the PAC dosage is iteratively adjusted until the water quality requirements are met;

[0013] S5: Model deployment and real-time control: Based on the optimized multi-stage model, the water quality data from the SCADA system is received in real time, and the dosage of NaClO, FeCl3 and PAC is predicted and dynamically adjusted in turn to output the optimal dosing plan.

[0014] The real-time raw water quality parameters, flow rate, dosage data, daily water quality data and effluent turbidity in the present invention are derived from online sensors and manual detection record data of the water purification process.

[0015] The present invention is the first to use the pre-chlorination NaClO dosage as the characteristic input of the coagulant dosage prediction model, and systematically quantify the nonlinear coupling effect of pre-chlorination on coagulant dosage.

[0016] Furthermore, the dynamic data management in S2 specifically includes: resampling the hourly online monitoring water quality data and dosage data into daily data based on the median, merging them with the daily water quality data, using forward filling and linear interpolation to supplement the low-frequency detection water quality parameters, and using dynamic box plot method and sliding mean filtering to replace abnormal values; the window length of the sliding mean filter is N=5.

[0017] In the present invention, missing values ​​are processed by combining forward filling and linear interpolation, dynamic box plot method is used for adaptive processing of outliers, and sliding mean filtering (window length N=5) is performed to replace outliers with window mean.

[0018] This invention enhances the prediction robustness of the model through multi-source data fusion and dynamic feature engineering. It integrates the SCADA system, integrated information management system and manual daily inspection data through the hourly-daily data resampling method to unify data accuracy and provide new ideas for innovative data island management.

[0019] Furthermore, the multi-stage prediction framework of pre-chlorination NaClO→FeCl3→PAC→effluent turbidity in S3 is specifically as follows:

[0020] S3.1: First, establish a NaClO dosage prediction model for pre-chlorination, using the XGBoost regression model. The model input is the raw water quality parameters in S1, and grid search is used to adjust the parameters.

[0021] S3.2: Secondly, a FeCl3 dosage prediction model is established using the XGBoost regression model. The model inputs are the raw water quality parameters and pre-chlorination NaClO dosage in S1, and grid search is used to adjust the parameters.

[0022] S3.3: The third step is to establish a PAC dosage prediction model and build a classification-regression coupling mechanism. Based on the intermittent dosage characteristics of PAC in winter, a support vector machine classification model is designed to predict the dosage status. If dosage is required, the XGBoost regression model is used to predict the PAC dosage.

[0023] S3.4: Finally, the effluent turbidity prediction model is established. The model input is the raw water quality parameters in S1, the pre-chlorination NaClO dosage, the FeCl3 dosage and the PAC dosage, and the grid search is used to adjust the parameters.

[0024] This paper proposes for the first time a multi-stage prediction model framework of pre-chlorination NaClO→FeCl3→PAC→effluent turbidity. Compared with the research that only uses water quality parameters as model input, it establishes a cross-domain feature association of "water quality-reagent-process".

[0025] Furthermore, the classification-regression coupling mechanism in S3.3 is specifically as follows:

[0026] S3.3.1: First, design a classification algorithm to determine whether PAC is added. Mark the PAC dosage data as "0" or "1" to indicate non-addition and addition, respectively. Combine the water quality data and previous drug addition data to form a classification data set.

[0027] S3.3.2: Construct a support vector machine classification model based on the classification data set, using grid search to adjust the parameters. The C regularization parameter ranges from 0.5 to 10 with a step size of 0.5, and the coefficient of the gamma kernel function ranges from 0.01 to 1 with a step size of 0.01. Determine the PAC dosage status and output "0" or "1".

[0028] S3.3.3: If the classification result is "1", a PAC dosage prediction model is established using the XGBoost regression model. The model inputs are the raw water quality parameters in S1, the pre-chlorination NaClO dosage, and the FeCl3 dosage. Grid search is used to adjust the parameters.

[0029] To address the problem of PAC deactivation in winter, the present invention establishes a prediction structure that first classifies and then regresses. The dosage demand is predicted by the SVM classifier, and the dosage is predicted by XGBoost regression. This breaks through the limitations of the traditional single regression model and significantly improves the accuracy of drug consumption control. It is particularly suitable for water purification processes of source water.

[0030] Furthermore, the grid search parameters are set as follows: n_estimators ranges from 50 to 150 with a step size of 5, max_depth ranges from 3 to 10 with a step size of 1, and learning_rate ranges from 0.01 to 0.1 with a step size of 0.01.

[0031] Furthermore, the adjustment strategy of the dynamic feedback optimization in S4 is based on the SHAP feature importance analysis, which prioritizes the optimization of the PAC dosage, followed by the FeCl3 and pre-chlorination NaClO dosages.

[0032] The present invention combines SHAP interpretability analysis to dynamically screen key parameters, helping to deepen the understanding of the impact of environmental parameters on the coagulation process. The model is lightweight, has low deployment cost and is easy to implement, significantly enhances effluent water quality assurance, and can effectively reduce drug consumption and corresponding carbon emissions.

[0033] Furthermore, the threshold value of the effluent turbidity in S4 is 0.3NTU.

[0034] Furthermore, the output of the effluent turbidity prediction model in S3.4 is used for closed-loop feedback control.

[0035] Furthermore, the real-time raw water quality parameters in S1 include raw water turbidity, pH, temperature, ammonia nitrogen, free chlorine and total chlorine;

[0036] The dosage data includes pre-chlorination NaClO dosage, FeCl3 dosage and PAC dosage;

[0037] The daily water quality data includes temperature at multiple points, total alkalinity, chloride, protein nitrogen, nitrite nitrogen, oxygen consumption, total hardness, residual aluminum, chlorophyll, chlorophyll a, algae count and total colony count.

[0038] On the other hand, the present invention provides the application of any of the above-mentioned intelligent low-carbon coagulation dosing methods based on multi-stage machine learning and dynamic feedback in the multi-agent coordinated control scenario of a water source water plant.

[0039] Beneficial effects:

[0040] (1) The present invention constructs a multi-stage prediction architecture of pre-chlorination NaClO→FeCl3→PAC→effluent turbidity chain prediction logic that adapts to the characteristics of water source water, and corrects subsequent prediction values ​​through feedback of the previous reagent dosage to suppress error transmission; at the same time, based on the intermittent addition characteristics of PAC, a dynamic classification and regression mechanism of classification judgment + regression prediction is established; an efficient data governance method for multi-system and multi-precision data of water plants is designed to solve the outlier processing and feature engineering optimization in the mixed scenario of minute-level and daily precision data; by comparing and optimizing machine learning models, a lightweight model based on XGBoost is developed, which significantly improves the adaptability to water source water and the robustness of the model.

[0041] (2) The present invention fills the gaps in the existing technologies in multi-agent synergistic prediction, dynamic adaptation of intermittent dosing, and low-quality data modeling, providing water source and water plants with a high-precision, explainable, and easy-to-deploy intelligent precision dosing solution, thereby promoting the transformation of the water treatment industry towards intelligent, efficient, low-consumption, and low-carbon development.

[0042] (3) The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback provided by the present invention takes the pre-chlorination NaClO dosage as the feature input of the coagulant dosage prediction model for the first time, and systematically quantifies the nonlinear coupling effect of pre-chlorination on coagulant dosage; it proposes for the first time a multi-stage prediction model framework of pre-chlorination NaClO→FeCl3→PAC→effluent turbidity. Compared with the research that only uses water quality parameters as model input, it establishes a cross-domain feature association of water quality-agent-process; in response to the problem of PAC decommissioning in winter, a prediction structure of first classification and then regression is established, and the dosage demand is predicted by the SVM classifier, and the dosage is predicted by XGBoost regression, which breaks through the limitations of the traditional single regression model and significantly improves the drug consumption control accuracy, which is particularly suitable for water purification processes of source water;

[0043] (4) The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback provided by the present invention enhances the prediction robustness of the model through multi-source data fusion and dynamic feature engineering, integrates the SCADA system, the integrated information management system and the manual daily inspection data through the hourly-daily data resampling method, unifies the data accuracy, and provides new ideas for innovative data island management; and combines SHAP interpretability analysis to dynamically screen key parameters, helping to deepen the understanding of the influence of environmental parameters on the coagulation process.

[0044] (5) In the coagulation dosing method provided by the present invention, the model is lightweight, has low deployment cost and is easy to implement, significantly enhances the effluent water quality assurance, and can effectively reduce drug consumption and corresponding carbon emissions. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic flow chart of the coagulation dosing method of the present invention.

[0046] Figure 2-Figure 5 The performance graphs of Pre-NaClO (pre-chlorinated NaClO), FeCl3, PAC (polyaluminum chloride) and Tur-eff (effluent turbidity) using the LINEAR (linear regression) model, XGBoost model and LSTM model on the test set in Example 1 of the present invention and Comparative Examples 1-2 are respectively.

[0047] Figure 6-Figure 9 They are the feature importance diagrams of the SHAP-based pre-chlorination NaClO, FeCl3, PAC (polyaluminium chloride) and effluent turbidity prediction model in Example 1 of the present invention. DETAILED DESCRIPTION

[0048] The present invention will be described below in conjunction with specific embodiments. It should be noted that the following examples are illustrative of the present invention and are intended only to illustrate the present invention and are not intended to limit the present invention. Other combinations and various modifications within the scope of the present invention may be made without departing from the spirit or scope of the present invention.

[0049] In the present invention, FeCl3 is ferric chloride, PAC is polyaluminium chloride, and NaClO is sodium hypochlorite, all of which are conventional reagents in the art.

[0050] Example 1

[0051] An intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback is constructed through the following steps:

[0052] S1: Obtain inlet and outlet water and process water quality indicators, flow rate, and hourly dosing data from the water source plant SCADA system. The data source is the online sensors of the water purification process. The main water quality monitoring indicators include turbidity, pH, temperature, and ammonia nitrogen at the outlet of the pre-sedimentation tank; turbidity, pH, temperature, and total chlorine of the water before filtration; and turbidity, pH, ammonia nitrogen, and total chlorine of the water leaving the plant. Dosing data includes the dosage of ferric chloride (FeCl3), polyaluminum chloride (PAC), and pre-chlorination sodium hypochlorite (NaClO);

[0053] S2: Obtain hourly dosing data and daily water quality data from the water plant's integrated information management system. Data sources include online sensors and manual inspection records. Daily water quality data includes temperature, total alkalinity, chloride, protein nitrogen, nitrite nitrogen, oxygen consumption, total hardness, residual aluminum, chlorophyll, chlorophyll a, algae count, and total colony count at multiple points.

[0054] S3: Convert the hourly dosage data into units to obtain the dosage data of pre-chlorination NaClO, FeCl3 and PAC (in terms of stock solution, mg / L);

[0055] S4: To increase the multi-dimensional water quality information and take into account the smooth transition of the actual dosage of water plants, the hourly data were resampled into daily data with the median as the representative;

[0056] S5: Combine the above daily data with the daily water quality data to obtain the following modeling input data: turbidity (Tur, NTU), pH, ammonia nitrogen (AN, mg / L), temperature (Tem, °C), total alkalinity (Alka, mg / LCaCO3), chlorides (Chlorides, mg / L), protein nitrogen (PN, mg / L), oxygen consumption (OD, mg / L), total hardness (Hard, mg / L CaCO3), chlorophyll (Chlorophyll, μg / L), chlorophyll-a (Chlorophyll-a, μg / L), algae count (Algae, 10,000 / L), total colony count (Colony, CFU / mL), pre-chlorination NaClO dosage (Pre-NaClO, mg / L), FeCl3 dosage (mg / L), PAC dosage (mg / L), and factory water turbidity (Tur-eff, NTU) from August 1, 2017 to July 31, 2022.

[0057] S6: Missing values ​​were handled using a combination of forward filling and linear interpolation, and outlier adaptive processing was performed using a dynamic boxplot method. Sliding mean filtering (window length N = 5) was performed to replace outliers with the window mean. After data preprocessing, the data was divided into training and test sets, with 1528 data in the training set and 200 data in the test set.

[0058] S7: Construct a multi-stage prediction framework, and sequentially establish the dosage prediction models of pre-chlorination NaClO, FeCl3 and PAC, as well as the effluent turbidity prediction model. The model inputs are water quality data and the dosage data of the previous reagents.

[0059] S8: First, establish a pre-chlorination NaClO dosage prediction model using the XGBoost regression model. The model input is the raw water quality parameters in S5. Grid search is used to adjust the parameters. The value range of n_estimators is 50-150 (step size is 5), the value range of max_depth is 3-10 (step size is 1), and the value range of learning_rate is 0.01-0.1 (step size is 0.01).

[0060] S9: Secondly, a FeCl3 dosage prediction model is established using the XGBoost regression model. The model inputs are the raw water quality parameters and pre-chlorination NaClO dosage in S5. The grid search parameter settings are the same as S8.

[0061] S10: The third step is to establish a PAC dosage prediction model, which specifically includes the following steps:

[0062] S10.1: Due to the intermittent deactivation of PAC in winter, a classification algorithm is first designed to determine whether PAC is added. The PAC dosage data are marked as "0" and "1" to indicate non-dosage and addition, respectively. The classification data is then combined with the water quality data and the previous dosing data to form a classification data set.

[0063] S10.2: Construct a support vector machine (SVM) classification model based on the classification data set. Grid search is used to adjust the parameters. The C regularization parameter ranges from 0.5 to 10 (with a step size of 0.5) and the gamma kernel function coefficient ranges from 0.01 to 1 (with a step size of 0.01). The model determines the PAC dosing status and outputs "0" (not dosing) or "1" (dosing required).

[0064] S10.3: If the classification result is "1", the XGBoost regression model is used. The model input is the raw water quality parameters in S5, the pre-chlorination NaClO dosage, and the FeCl3 dosage. The grid search parameter settings are the same as S8.

[0065] S11: Finally, establish the effluent turbidity prediction model. The model input is the raw water quality parameters in S5, the pre-chlorination NaClO dosage, the FeCl3 dosage and the PAC dosage. The grid search parameter settings are the same as S8.

[0066] Comparative Example 1

[0067] The same as Example 1, except that the XGBoost model is replaced by a linear regression (LINEAR) model.

[0068] Comparative Example 2

[0069] The same as Example 1, except that the XGBoost model is changed to an LSTM model. The optimized hyperparameters of the LSTM model include hidden_size, num_layers, batch_size, and seq_len.

[0070] Performance Testing

[0071] 1. The regression prediction models of Example 1 and Comparative Examples 1-2 were trained and tested, and MAPE, MAE, and RMSE were used as model performance evaluation indicators. The performance evaluation results of the regression prediction models are shown in Tables 1 and Figure 2-Figure 5 As shown:

[0072] Table 1

[0073]

[0074] According to Table 1 and Figure 2-5 From the comparison of performance results, it can be seen that the prediction accuracy of the XGBoost model for pre-chlorination NaClO, FeCl3, PAC and effluent turbidity using the present invention is the best in terms of MAPE, MAE and RMSE model evaluation parameters. At the same time, the prediction error (MAPE) of the effluent turbidity of the model of the present invention is reduced from 21.9% to 5.5% compared with the prior art, which can significantly improve the prediction accuracy.

[0075] 2. Perform feature importance analysis of the model in Example 1 using SHAP, as shown in Figure 6-Figure 9 shown.

[0076] according to Figure 6-Figure 9 The results analysis shows that the main factors affecting the accuracy of the pre-chlorination NaClO prediction model are temperature, turbidity and total hardness; the main factors affecting the accuracy of the FeCl3 prediction model are turbidity, alkalinity and temperature; the main factors affecting the accuracy of the PAC prediction model are temperature, turbidity and pre-chlorination NaClO dosage; the main factors affecting the accuracy of the effluent turbidity prediction model are PAC dosage, temperature and total bacterial count in raw water.

[0077] Based on the above information, when it is determined that the predicted effluent turbidity exceeds the threshold (>0.3NTU), the PAC dosage is adjusted iteratively until the water quality requirements are met. At the same time, it can assist the front-line operators of the water plant to implement on-site control based on key influencing parameters.

[0078] The above embodiments are only for illustrating the technical concept and features of the present invention. Its purpose is to enable people familiar with this technology to understand the content of the present invention and implement it. It cannot be used to limit the scope of protection of the present invention. Any equivalent changes or modifications made according to the spirit of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback, characterized in that: The following steps are involved: S1: Multi-source data acquisition and fusion: Obtain real-time raw water quality parameters, flow rate, dosage data, daily water quality data and effluent turbidity from the water plant SCADA system and integrated information management system; S2: Dynamic data management: Hourly data is converted to daily data, the median is resampled, and the data is merged with the daily inspection data. Missing values ​​and outliers are processed. S3: Multi-stage prediction framework construction, specifically: S3.1: First, establish a NaClO dosage prediction model for pre-chlorination, using the XGBoost regression model. The model input is the raw water quality parameters in S1, and grid search is used to adjust the parameters. S3.2: Secondly, a FeCl3 dosage prediction model is established using the XGBoost regression model. The model inputs are the raw water quality parameters and pre-chlorination NaClO dosage in S1, and grid search is used to adjust the parameters. S3.3: The third step is to establish a PAC dosage prediction model and build a classification-regression coupling mechanism. Based on the intermittent dosage characteristics of PAC in winter, a support vector machine classification model is designed to predict the dosage status. If dosage is required, the XGBoost regression model is used to predict the PAC dosage. S3.4: Finally, establish the effluent turbidity prediction model. The model input is the raw water quality parameters in S1, the pre-chlorination NaClO dosage, the FeCl3 dosage and the PAC dosage. The grid search is used to adjust the parameters. S4: Dynamic feedback optimization: Based on the effluent turbidity prediction results, if the predicted value exceeds the threshold, the PAC dosage is iteratively adjusted until the water quality requirements are met; S5: Model deployment and real-time control: Based on the optimized multi-stage model, the water quality data from the SCADA system is received in real time, and the dosage of NaClO, FeCl3 and PAC is predicted and dynamically adjusted in turn to output the optimal dosing plan.

2. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 1 is characterized in that: The dynamic data management in S2 specifically includes: resampling the hourly online monitoring water quality data and dosage data into daily data based on the median, merging them with the daily water quality data, using forward filling and linear interpolation to supplement the low-frequency detection water quality parameters, and using dynamic box plot method and sliding mean filtering to replace abnormal values; the window length of the sliding mean filter is N=5.

3. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 1 is characterized in that: The classification-regression coupling mechanism in S3.3 is specifically as follows: S3.3.1: First, design a classification algorithm to determine whether PAC has been added. Mark the PAC dosage data as "0" or "1" to indicate non-addition and addition, respectively. Combine the data with the water quality data and the previous drug addition data to form a classification data set. S3.3.2: Construct a support vector machine classification model based on the classification data set. Use grid search to adjust the parameters. The C regularization parameter ranges from 0.5 to 10 with a step size of 0.

5. The coefficient of the gamma kernel function ranges from 0.01 to 1 with a step size of 0.

01. Determine the PAC dosage status and output "0" or "1". S3.3.3: If the classification result is "1", a PAC dosage prediction model is established using the XGBoost regression model. The model inputs are the raw water quality parameters in S1, the pre-chlorination NaClO dosage, and the FeCl3 dosage. Grid search is used to adjust the parameters.

4. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 3 is characterized in that: The grid search parameters are set as follows: n_estimators ranges from 50 to 150 with a step size of 5, max_depth ranges from 3 to 10 with a step size of 1, and learning_rate ranges from 0.01 to 0.1 with a step size of 0.

01.

5. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 1 is characterized in that: The adjustment strategy of the dynamic feedback optimization in S4 is based on the SHAP feature importance analysis, which prioritizes the optimization of the PAC dosage, followed by the FeCl3 and pre-chlorination NaClO dosages.

6. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 1 is characterized in that: The threshold value of the effluent turbidity in S4 is 0.3 NTU.

7. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 3 is characterized in that: The output of the effluent turbidity prediction model in S3.4 is used for closed-loop feedback control.

8. The intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback according to claim 1 is characterized in that: The real-time raw water quality parameters in S1 include raw water turbidity, pH, temperature, ammonia nitrogen, free chlorine and total chlorine; The dosage data includes pre-chlorination NaClO dosage, FeCl3 dosage and PAC dosage; The daily water quality data includes temperature at multiple points, total alkalinity, chloride, protein nitrogen, nitrite nitrogen, oxygen consumption, total hardness, residual aluminum, chlorophyll, chlorophyll a, algae count and total colony count.

9. Application of the intelligent low-carbon coagulation dosing method based on multi-stage machine learning and dynamic feedback as described in any one of claims 1-8 in the multi-agent coordinated control scenario of a water source water plant.

Images