Non-point source pollution risk assessment method based on machine learning

Abstract

The invention relates to the technical field of pollution risk assessment, and particularly discloses a non-point source pollution risk assessment method based on machine learning, and the method comprises the steps: collecting the water quality monitoring data of a target drainage basin under various meteorological conditions in a historical period, and synchronously collecting the meteorological monitoring data of a corresponding period; the method comprises the steps that meteorological extreme events and water quality response characteristics corresponding to the meteorological extreme events are extracted, a meteorological-water quality associated data set is constructed, a prediction model between meteorological factors and pollution load peak values is trained through a machine learning method, and the short-term risk level of non-point source pollution is evaluated based on a prediction result; and generating a risk early warning signal according to the risk level obtained by evaluation, and through multi-source data fusion and machine learning modeling, providing high-timeliness and high-precision risk decision support for drainage basin management, significantly enhancing the ability to cope with sudden pollution events, and improving the risk early warning effect. And meanwhile, key technical guarantee is provided for optimizing and treating resource allocation and reducing water environment disaster risks.

Description

Technical Field

[0001] This invention belongs to the field of pollution risk assessment technology, and relates to a non-point source pollution risk assessment method based on machine learning. Background Technology

[0002] The effectiveness of non-point source pollution control and treatment largely depends on the ability to accurately capture and predict its risks, especially short-term high-risk events. Utilizing the correlation between extreme meteorological events (such as short-duration heavy rainfall) and water quality data to dynamically capture and predict sudden high-risk pollution loads is of paramount importance and necessity. This is because non-point source pollution has a typical "event-based" characteristic, with the vast majority of its load concentrated in a few extreme rainfall events. Conventional risk assessment methods based on long-term averages often smooth out these peak events, leading to a significant underestimation of their true environmental impact.

[0003] However, current mainstream technologies still have significant technical shortcomings and drawbacks in this regard:

[0004] (1) At the data level, the existing monitoring system is often “hydrology-water quality” separated, and the data is not synchronized in time and space. The water quality sampling frequency is usually much lower than that of hydrological and meteorological monitoring, resulting in a huge blind spot when capturing the peak concentration of pollutants that completely corresponds to a specific rainstorm event, and failing to provide high-resolution “event-response” paired data for model construction and validation;

[0005] (2) At the model level, most of the widely used assessment models are static or quasi-static models, and their inputs are mostly annual or monthly average data, which are essentially unable to respond to the dynamic process of a single rainfall event. Even if a complex distributed hydrological model is used, its calibration and verification are usually based on monthly or annual loads, which means that although the model can simulate the long-term equilibrium state, it seriously underestimates the peak flow and the corresponding peak pollution load, resulting in a systematic bias in the prediction of extreme events. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention provides a non-point source pollution risk assessment method based on machine learning to solve the above-mentioned technical problems.

[0007] To achieve the above and other objectives, the technical solution adopted by the present invention is as follows:

[0008] This invention provides a non-point source pollution risk assessment method based on machine learning, the method comprising:

[0009] Step S1: Collect water quality monitoring data of the target watershed under various meteorological conditions during historical periods, and simultaneously collect meteorological monitoring data for the corresponding periods;

[0010] Step S2: Based on the collected water quality monitoring data and meteorological monitoring data, extract meteorological extreme events and their corresponding water quality response features, and construct a meteorological-water quality correlation dataset. The meteorological extreme events include short-term heavy precipitation events and continuous rainfall events.

[0011] Step S3: Use machine learning methods to train a prediction model between meteorological factors and pollution load peaks, and assess the short-term risk level of non-point source pollution based on the prediction results.

[0012] Step S4: Based on the risk level obtained from the assessment, generate a risk warning signal. If the risk level exceeds the preset threshold, trigger the warning and output the corresponding risk prevention and control suggestions.

[0013] As described above, the non-point source pollution risk assessment method based on machine learning provided by the present invention has at least the following beneficial effects:

[0014] (1) This invention constructs a highly timely basic database by collecting water quality monitoring data and synchronous meteorological monitoring data of the target watershed under various meteorological conditions during historical periods; accurately extracts meteorological extreme events such as short-term heavy precipitation events and continuous rainfall events and their corresponding water quality response characteristics based on the collected data to form a meteorological-water quality correlation dataset; uses machine learning methods to train a prediction model between meteorological factors and pollution load peaks to achieve accurate assessment of the short-term risk level of non-point source pollution; finally generates a risk warning signal based on the assessment results, and automatically triggers the warning and outputs risk prevention and control suggestions when the risk exceeds the preset threshold. This invention not only achieves accurate capture, dynamic assessment and short-term warning of high-risk events of non-point source pollution in terms of technology, but also effectively solves the problems of delayed response to sudden pollution loads and static risk assessment results in traditional assessment methods at the application level, greatly improving the initiative and foresight of risk management.

[0015] (2) This invention provides timely and accurate risk decision support for watershed management through multi-source data fusion and machine learning modeling, significantly enhancing the ability to respond to sudden pollution events. At the same time, it provides key technical support for optimizing the allocation of governance resources and reducing the risk of water environment disasters, effectively enhancing the ability to respond to sudden non-point source pollution events, and fully meeting the urgent needs of modern environmental management for intelligent, refined and real-time technologies. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram showing the connections between the steps of the method of the present invention. Detailed Implementation

[0018] The following description, in conjunction with the implementation of this invention, is merely an example and illustration of the concept of this invention. Those skilled in the art can make various modifications or additions to the specific embodiments described, or use similar methods to replace them, as long as they do not deviate from the inventive concept or exceed the scope defined in these claims, all of which should fall within the protection scope of this invention.

[0019] Example 1

[0020] Please see Figure 1 As shown, a non-point source pollution risk assessment method based on machine learning is proposed, which includes:

[0021] Step S1: Collect water quality monitoring data of the target watershed under various meteorological conditions during historical periods, and simultaneously collect meteorological monitoring data for the corresponding periods;

[0022] The water quality monitoring data includes instantaneous pollution load values ​​at each sampling time, and the meteorological monitoring data includes precipitation, rainfall intensity, temperature, and humidity at each sampling time; the specific data collection process is as follows:

[0023] Based on historical meteorological data of the target watershed, meteorological scenario types were classified, resulting in four typical meteorological scenario classifications: clear weather, ordinary rainfall, short-term heavy rainfall, and continuous heavy rainfall.

[0024] Based on meteorological scenario classification data, differentiated water quality sampling strategies were developed. In the case of short-term heavy rainfall and continuous heavy rainfall, a high-frequency sampling mode was activated, while a conventional sampling mode was used in ordinary rainfall scenarios and a baseline sampling mode was used in sunny weather scenarios, generating water quality sampling scheme data for multiple scenarios.

[0025] Based on the data from the multi-scenario water quality sampling scheme, water quality monitoring data is collected by online water quality monitoring stations deployed at the watershed cross section, and meteorological monitoring data is simultaneously acquired through automatic meteorological data acquisition equipment. The water quality monitoring data includes the concentration of pollutants and instantaneous flow data at each sampling time, and the meteorological monitoring data includes precipitation, rainfall intensity, temperature and humidity at each sampling time, forming the original synchronous monitoring dataset.

[0026] The original synchronous monitoring dataset was time-aligned and missing value imputation was performed to ensure that the timestamps of water quality and meteorological data were strictly matched, resulting in a time-aligned watershed synchronous monitoring database.

[0027] Based on a time-aligned watershed synchronous monitoring database, the concentration of pollutants and the corresponding instantaneous flow data at each sampling time are extracted. By multiplying the concentration of pollutants and the instantaneous flow data, the instantaneous value of the pollution load at each sampling time is calculated.

[0028] In this embodiment of the invention, the process of collecting and processing water quality and meteorological synchronous monitoring data of the target watershed is as follows:

[0029] First, based on historical meteorological data, a statistical distribution analysis method was used to classify meteorological scenarios. By calculating the percentile thresholds of multi-year precipitation data, meteorological conditions were categorized into four typical types: clear weather, ordinary rainfall, short-duration heavy rainfall, and continuous heavy rainfall, forming meteorological scenario classification data. Then, based on this classification, differentiated water quality sampling strategies were developed. A high-frequency sampling mode was activated for short-duration heavy rainfall and continuous heavy rainfall scenarios, with a sampling interval of no more than 15 minutes. A conventional sampling mode of hourly sampling was used for ordinary rainfall scenarios, and a baseline sampling mode of once every 4 hours was adopted for clear weather, thus generating a multi-scenario water quality sampling scheme with scenario adaptability. Subsequently, according to the above scheme, water quality parameters were collected in real time through online water quality monitoring stations deployed at key sections of the watershed. The data are obtained synchronously from meteorological elements acquired by automatic weather stations. Water quality monitoring data includes the concentrations of pollutants such as chemical oxygen demand, total nitrogen, and total phosphorus at each sampling time, as well as the instantaneous flow rate at the cross section. Meteorological data includes precipitation, rainfall intensity, temperature, and humidity at the same time point, which together constitute the original synchronous monitoring dataset. The original dataset is time-aligned and cleaned. The missing time periods are repaired using the sliding window interpolation method, and the timestamp matching algorithm is used to ensure strict synchronization between meteorological and water quality data, forming a time-series aligned watershed synchronous monitoring database. Finally, based on this time-series database, the pollutant concentration and instantaneous flow rate at each sampling time are extracted. The pollutant concentration value at each time point is multiplied by the corresponding instantaneous flow rate value to calculate the instantaneous pollution load value at each sampling point.

[0030] Step S2: Based on the collected water quality monitoring data and meteorological monitoring data, extract meteorological extreme events and their corresponding water quality response features, and construct a meteorological-water quality correlation dataset. The meteorological extreme events include short-term heavy precipitation events and continuous rainfall events.

[0031] Step S2 includes:

[0032] Step S21: Determine the extreme event threshold for precipitation and rainfall intensity in the meteorological monitoring data to obtain the meteorological extreme event identification results for short-term heavy precipitation events and continuous rainfall events. The identification results include the start time, end time, duration and maximum rainfall intensity of each meteorological event.

[0033] Step S22: Based on the time range in the meteorological extreme event identification results, extract the pollution factor concentration sequence and flow monitoring data within the corresponding time period from the water quality monitoring data to obtain the event-related water quality raw dataset;

[0034] Step S23: Perform time-series fluctuation analysis on the concentration sequence of pollutants in the original water quality dataset associated with the event, extract the peak concentration, concentration rise rate and high load duration of each pollutant, and generate a set of water quality response characteristic indicators;

[0035] Step S24: Associate and map the meteorological extreme event identification results with the water quality response characteristic index set according to time labels, and integrate them to form a meteorological-water quality association dataset.

[0036] In this embodiment of the invention, firstly, based on the precipitation and rainfall intensity sequences in meteorological monitoring data, a percentile statistical method is used to determine the threshold standard for extreme rainfall events. Short-duration heavy rainfall events and continuous rainfall events are identified by scanning a sliding time window, and the start time, end time, duration, and maximum rainfall intensity within each meteorological event are recorded to form a structured meteorological extreme event identification result. Then, based on the time stamps of each event in the identification result, the concentration sequences of all pollutants within the corresponding time period and the synchronously monitored flow data are extracted from the water quality monitoring database to construct an event-related water quality raw dataset. The dataset is then subjected to temporal fluctuation feature analysis, and the peak concentration of each pollutant during the event process is extracted using a local extremum detection algorithm. The concentration rise rate is calculated using the ratio of the concentration difference between adjacent time points to the time interval, and the duration of the high pollution load state is determined based on a set load threshold, thereby generating a water quality response feature index set containing multi-dimensional features. Finally, the meteorological extreme event identification result and the water quality response feature index set are associated and matched according to time labels, and integrated into a structured meteorological-water quality associated dataset through database connection operations.

[0037] Step S22 includes the following sub-steps:

[0038] Step S221: Based on the start and end times of each event in the meteorological extreme event identification results, determine the target extraction time interval for water quality monitoring data and generate time window identifier data corresponding to the event;

[0039] Step S222: Based on the time window identifier data, extract the original sequence of pollutant concentrations within each time window and the synchronously monitored flow data from the water quality monitoring database to obtain the water quality data associated with the initial screening event;

[0040] Step S223: Perform integrity verification and outlier removal on the pollutant concentration sequence in the water quality data associated with the initial screening events, and repair the missing data by time adjacent interpolation to form the repaired event water quality dataset;

[0041] Step S224: Align the pollutant concentration sequence in the restored event water quality dataset with the flow monitoring data according to the sampling time, and integrate them to generate the event-related water quality original dataset.

[0042] In this embodiment of the invention, based on the start and end times of each event in the meteorological extreme event identification results, a timestamp parsing algorithm is used to determine the data extraction time window corresponding to each event. A time buffering expansion strategy is adopted to extend the event window forward and backward by a set duration to fully capture the hydrological response process, generating event time window identification data containing event number and time range marker. Based on this time window identification data, the original concentration sequences of all pollutants within each time window and the instantaneous flow data monitored simultaneously are retrieved from the water quality monitoring database to form preliminary event-related water quality data. The preliminary screening data is quality-verified, and abnormal concentration values ​​exceeding the reasonable range are identified and removed using the standard deviation threshold method. The missing data period is determined by time series continuity detection, and the missing data is repaired based on the linear interpolation method of adjacent time point monitoring values ​​to ensure data continuity and completeness, forming a repaired event water quality dataset. Finally, the repaired pollutant concentration sequence and flow monitoring data are precisely aligned with millisecond-level timestamps and integrated into a structured and complete event-related water quality original dataset using a time series merging algorithm.

[0043] Step S23 includes:

[0044] Step S231: Based on the event-related water quality raw dataset, perform local extreme point scanning on the concentration sequence of each pollutant, identify the peak points in the concentration sequence and the corresponding sampling time, and generate pollutant peak point data;

[0045] Step S232: Based on the sampling time in the peak point data of pollutants, trace back to the initial rise time of the concentration sequence, calculate the ratio of the difference between the peak point concentration and the initial point concentration to the time interval, and obtain the pollutant concentration rise rate data.

[0046] Step S233: Based on the peak concentration in the peak location data of pollutants, set a high load concentration threshold, determine the continuous time period in the concentration sequence that exceeds the threshold, calculate the duration of the time period, and obtain the high load duration data of pollutants.

[0047] Step S234: Integrate the peak location data of pollutants, the concentration rise rate data, and the high load duration data according to the type of pollutant to generate a set of water quality response characteristic indicators.

[0048] In this embodiment of the invention, the specific implementation process of step S23 is as follows: Based on the original event-related water quality dataset, the sliding window extreme value detection algorithm is first used to scan the concentration sequence of each pollutant. By comparing the numerical relationship between adjacent data points and data in the preceding and following windows, local peak points are identified. The concentration value and precise sampling time of each peak point are recorded to generate peak point data containing pollutant type, peak concentration, and occurrence time. Based on the sampling time in the peak point data, the starting time of the current rise in the concentration sequence is traced back. The concentration change per unit time is obtained by calculating the difference between the peak point concentration and the starting point concentration and dividing by the time interval between the two, which is used as the concentration rise rate data of the pollutant. Based on the peak concentration value in the peak point data, a high load concentration threshold is set in combination with historical water quality standards or watershed background values. The time period in the concentration sequence that continuously exceeds the threshold is scanned. The duration of the high load state is calculated by the difference between the end time and the start time of the time period. Finally, the peak point data, concentration rise rate data, and high load duration data of each pollutant are structurally integrated according to the pollutant type to form a water quality response characteristic index set containing multi-dimensional characteristic indicators.

[0049] Step S3: Use machine learning methods to train a prediction model between meteorological factors and pollution load peaks, and assess the short-term risk level of non-point source pollution based on the prediction results.

[0050] It should be noted that the meteorological factors mentioned in this invention refer to precipitation, rainfall intensity, temperature, and humidity.

[0051] Step S3 includes:

[0052] Step S31: Based on the meteorological-water quality correlation dataset, extract meteorological feature parameters as model input features, and extract pollution load peak data as prediction labels to construct a machine learning sample set;

[0053] Step S32: Standardize the input features in the machine learning sample set, and divide the sample set into a training set and a test set in chronological order to obtain the standardized training sample set and test sample set;

[0054] Step S33: Select random forest as the prediction model, train the model using the training sample set, optimize the model hyperparameters through cross-validation, and obtain the trained pollution load prediction model.

[0055] Step S34: Based on the pollution load prediction model, predict the peak pollution load of real-time meteorological data, compare the predicted value with the preset risk threshold, and determine the short-term risk level of non-point source pollution.

[0056] In this embodiment of the invention, the specific implementation process of step S3 is as follows: First, based on the meteorological-water quality correlation dataset, precipitation, rainfall intensity, temperature, and humidity are selected as key meteorological feature parameters, and pollution load peak data are extracted as prediction labels. A machine learning sample set is constructed through feature-label pairing. The input features in the sample set are Z-score standardized to make their mean zero and standard deviation one. The sample set is divided into a training set and a test set according to the time series order to ensure that the training data is earlier than the test data to avoid future information leakage. Random forest is selected as the prediction model. The training set is used to optimize hyperparameters such as the number of decision trees, maximum depth, and node splitting criteria through grid search. Five-fold cross-validation is used to evaluate the performance of each parameter combination and select the optimal configuration to train the pollution load prediction model. Finally, based on the model, the pollution load peak is predicted from real-time meteorological monitoring data. The predicted value is compared with the risk threshold set according to the statistical data of historical pollution events. The short-term risk level of non-point source pollution is determined according to the threshold range in which the predicted value is located, and a risk warning report is generated.

[0057] Step S33 includes the following sub-steps:

[0058] Initialize the random forest regression model, set the range of hyperparameter candidates for the number of decision trees, maximum depth, and feature sampling ratio, and generate a set of hyperparameter candidate combinations;

[0059] Based on the training sample set, the K-fold cross-validation method is used to train the model on the candidate hyperparameter combination set to obtain the cross-validation performance results under each hyperparameter combination.

[0060] Based on the cross-validation performance results, the mean square error and coefficient of determination for each hyperparameter combination are calculated to generate hyperparameter performance evaluation data.

[0061] The optimal hyperparameter configuration is selected by choosing the combination of hyperparameters with the smallest mean square error and the highest coefficient of determination. The final random forest model is then trained using the full training sample set to obtain the trained pollution load prediction model.

[0062] In this embodiment of the invention, the specific implementation process of step S33 is as follows: First, initialize the random forest regression model, and set the trial range of the number of decision trees, the maximum depth range of the trees, and the candidate values ​​of the feature sampling ratio of each tree according to domain experience. A systematic set of candidate hyperparameter combinations is formed by generating a parameter grid. Based on the standardized training sample set, the K-fold cross-validation method is used to train and validate the model for each hyperparameter combination. The training set is divided into K mutually exclusive subsets, which are used as the validation set in turn, and the rest are used as the training set. The prediction results of K validations under each hyperparameter combination are obtained and summarized as the cross-validation performance results. According to the performance results, the mean squared error and the mean coefficient of determination of K-fold validation under each hyperparameter combination are calculated respectively, where the mean squared error is the average of the squares of the differences between the predicted values ​​and the true values ​​of each validation set; the mean coefficient of determination is a measure of the degree of variation of the explanatory variables of the model.

[0063] Generate hyperparameter performance evaluation data that includes multiple evaluation indicators; finally, select the combination with the smallest average mean square error and the highest average coefficient of determination from all hyperparameter combinations as the optimal hyperparameter configuration, and retrain the random forest model with this optimal configuration using the entire training sample set to obtain a trained pollution load prediction model.

[0064] Step S34 includes the following sub-steps:

[0065] Step S341: Obtain real-time meteorological monitoring data, perform the same standardization processing on the data as on the training sample set, and generate standardized real-time meteorological feature data;

[0066] Step S342: Input standardized real-time meteorological characteristic data into the trained pollution load prediction model to obtain the predicted peak pollution load value;

[0067] Step S343: Compare the predicted peak pollution load value with preset risk thresholds at various levels, including low risk threshold, medium risk threshold and high risk threshold;

[0068] Step S344: Determine the risk level based on the comparison results. When the predicted value is lower than the low risk threshold, it is determined to be a low risk level. When the predicted value is between the low risk threshold and the medium risk threshold, it is determined to be a medium risk level. When the predicted value is between the medium risk threshold and the high risk threshold, it is determined to be a high risk level. When the predicted value exceeds the high risk threshold, it is determined to be a severe risk level.

[0069] In this embodiment of the invention, the specific implementation process of step S34 is as follows: First, real-time meteorological monitoring data is acquired. Using the same standardized parameters as the training sample set, namely the feature mean and standard deviation calculated during the training phase, the real-time precipitation, rainfall intensity, temperature, and humidity data are converted into standardized real-time meteorological feature data through the Z-score standardization method. The standardized data is input into the pre-trained pollution load prediction model. Parallel prediction is performed through multiple decision trees within the model, and the results are integrated to output the pollution load peak prediction value. The predicted value is compared step by step with three risk thresholds set in advance based on historical pollution event statistics, including a low-risk threshold based on the watershed background load, a medium-risk threshold based on light pollution events, and a high-risk threshold based on heavy pollution events. Finally, the risk level is determined according to the comparison results: when the predicted value is lower than the low-risk threshold, it is determined to be a low-risk level; when the predicted value is between the low-risk and medium-risk thresholds, it is determined to be a medium-risk level; when the predicted value is between the medium-risk and high-risk thresholds, it is determined to be a high-risk level; and when the predicted value exceeds the high-risk threshold, it is determined to be a severe-risk level, thus forming the short-term risk classification result for non-point source pollution.

[0070] Step S4: Based on the risk level obtained from the assessment, generate a risk warning signal. If the risk level exceeds the preset threshold, trigger the warning and output the corresponding risk prevention and control suggestions.

[0071] It should be noted that the interval and threshold sizes are set for ease of comparison. The size of the threshold depends on the amount of sample data and the base number set by those skilled in the art for each set of sample data, as long as it does not affect the proportional relationship between the parameter and the quantized value. Furthermore, the above formulas are all dimensionless calculations, and the formulas are derived from software simulations using a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0072] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0074] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-point source pollution risk assessment method based on machine learning, characterized in that, include: Step S1: Collect water quality monitoring data of the target watershed under various meteorological conditions during historical periods, and simultaneously collect meteorological monitoring data for the corresponding periods; Step S2: Based on the collected water quality monitoring data and meteorological monitoring data, extract meteorological extreme events and their corresponding water quality response features, and construct a meteorological-water quality correlation dataset. The meteorological extreme events include short-term heavy precipitation events and continuous rainfall events. Step S3: Use machine learning methods to train a prediction model between meteorological factors and pollution load peaks, and assess the short-term risk level of non-point source pollution based on the prediction results. Step S4: Based on the risk level obtained from the assessment, generate a risk warning signal. If the risk level exceeds the preset threshold, trigger the warning and output the corresponding risk prevention and control suggestions.

2. The non-point source pollution risk assessment method based on machine learning according to claim 1, characterized in that, The water quality monitoring data includes instantaneous pollution load values ​​at each sampling time, and the meteorological monitoring data includes precipitation, rainfall intensity, temperature, and humidity at each sampling time; the specific data collection process is as follows: Based on historical meteorological data of the target watershed, meteorological scenario types were classified, resulting in four typical meteorological scenario classifications: clear weather, ordinary rainfall, short-term heavy rainfall, and continuous heavy rainfall. Based on meteorological scenario classification data, differentiated water quality sampling strategies were developed. In the case of short-term heavy rainfall and continuous heavy rainfall, a high-frequency sampling mode was activated, while a conventional sampling mode was used in ordinary rainfall scenarios and a baseline sampling mode was used in sunny weather scenarios, generating water quality sampling scheme data for multiple scenarios. Based on the data from the multi-scenario water quality sampling scheme, water quality monitoring data is collected by online water quality monitoring stations deployed at the watershed cross section, and meteorological monitoring data is simultaneously acquired through automatic meteorological data acquisition equipment. The water quality monitoring data includes the concentration of pollutants and instantaneous flow data at each sampling time, and the meteorological monitoring data includes precipitation, rainfall intensity, temperature and humidity at each sampling time, forming the original synchronous monitoring dataset. The original synchronous monitoring dataset was time-aligned and missing value imputation was performed to ensure that the timestamps of water quality and meteorological data were strictly matched, resulting in a time-aligned watershed synchronous monitoring database. Based on a time-aligned watershed synchronous monitoring database, the concentration of pollutants and the corresponding instantaneous flow data at each sampling time are extracted. By multiplying the concentration of pollutants and the instantaneous flow data, the instantaneous value of the pollution load at each sampling time is calculated.

3. The non-point source pollution risk assessment method based on machine learning according to claim 1, characterized in that, Step S2 includes: Step S21: Determine the extreme event threshold for precipitation and rainfall intensity in the meteorological monitoring data to obtain the meteorological extreme event identification results for short-term heavy precipitation events and continuous rainfall events. The identification results include the start time, end time, duration and maximum rainfall intensity of each meteorological event. Step S22: Based on the time range in the meteorological extreme event identification results, extract the pollution factor concentration sequence and flow monitoring data within the corresponding time period from the water quality monitoring data to obtain the event-related water quality raw dataset; Step S23: Perform time-series fluctuation analysis on the concentration sequence of pollutants in the original water quality dataset associated with the event, extract the peak concentration, concentration rise rate and high load duration of each pollutant, and generate a set of water quality response characteristic indicators; Step S24: Associate and map the meteorological extreme event identification results with the water quality response characteristic index set according to time labels, and integrate them to form a meteorological-water quality association dataset.

4. The non-point source pollution risk assessment method based on machine learning according to claim 3, characterized in that, Step S22 includes the following sub-steps: Step S221: Based on the start and end times of each event in the meteorological extreme event identification results, determine the target extraction time interval for water quality monitoring data and generate time window identifier data corresponding to the event; Step S222: Based on the time window identifier data, extract the original sequence of pollutant concentrations within each time window and the synchronously monitored flow data from the water quality monitoring database to obtain the water quality data associated with the initial screening event; Step S223: Perform integrity verification and outlier removal on the pollutant concentration sequence in the water quality data associated with the initial screening events, and repair the missing data by time adjacent interpolation to form the repaired event water quality dataset; Step S224: Align the pollutant concentration sequence in the restored event water quality dataset with the flow monitoring data according to the sampling time, and integrate them to generate the event-related water quality original dataset.

5. The non-point source pollution risk assessment method based on machine learning according to claim 3, characterized in that, Step S23 includes: Step S231: Based on the event-related water quality raw dataset, perform local extreme point scanning on the concentration sequence of each pollutant, identify the peak points in the concentration sequence and the corresponding sampling time, and generate pollutant peak point data; Step S232: Based on the sampling time in the peak point data of pollutants, trace back to the initial rise time of the concentration sequence, calculate the ratio of the difference between the peak point concentration and the initial point concentration to the time interval, and obtain the pollutant concentration rise rate data. Step S233: Based on the peak concentration in the peak location data of pollutants, set a high load concentration threshold, determine the continuous time period in the concentration sequence that exceeds the threshold, calculate the duration of the time period, and obtain the high load duration data of pollutants. Step S234: Integrate the peak location data of pollutants, the concentration rise rate data, and the high load duration data according to the type of pollutant to generate a set of water quality response characteristic indicators.

6. The non-point source pollution risk assessment method based on machine learning according to claim 1, characterized in that, Step S3 includes: Step S31: Based on the meteorological-water quality correlation dataset, extract meteorological feature parameters as model input features, and extract pollution load peak data as prediction labels to construct a machine learning sample set; Step S32: Standardize the input features in the machine learning sample set, and divide the sample set into a training set and a test set in chronological order to obtain the standardized training sample set and test sample set; Step S33: Select random forest as the prediction model, train the model using the training sample set, optimize the model hyperparameters through cross-validation, and obtain the trained pollution load prediction model. Step S34: Based on the pollution load prediction model, predict the peak pollution load of real-time meteorological data, compare the predicted value with the preset risk threshold, and determine the short-term risk level of non-point source pollution.

7. The non-point source pollution risk assessment method based on machine learning according to claim 6, characterized in that, Step S33 includes the following sub-steps: Initialize the random forest regression model, set the range of hyperparameter candidates for the number of decision trees, maximum depth, and feature sampling ratio, and generate a set of hyperparameter candidate combinations; Based on the training sample set, the K-fold cross-validation method is used to train the model on the candidate hyperparameter combination set to obtain the cross-validation performance results under each hyperparameter combination. Based on the cross-validation performance results, the mean square error and coefficient of determination for each hyperparameter combination are calculated to generate hyperparameter performance evaluation data. The optimal hyperparameter configuration is selected by choosing the combination of hyperparameters with the smallest mean square error and the highest coefficient of determination. The final random forest model is then trained using the full training sample set to obtain the trained pollution load prediction model.

8. The non-point source pollution risk assessment method based on machine learning according to claim 6, characterized in that, Step S34 includes the following sub-steps: Step S341: Obtain real-time meteorological monitoring data, perform the same standardization processing on the data as on the training sample set, and generate standardized real-time meteorological feature data; Step S342: Input standardized real-time meteorological characteristic data into the trained pollution load prediction model to obtain the predicted peak pollution load value; Step S343: Compare the predicted peak pollution load value with preset risk thresholds at various levels, including low risk threshold, medium risk threshold and high risk threshold; Step S344: Determine the risk level based on the comparison results. When the predicted value is lower than the low risk threshold, it is determined to be a low risk level. When the predicted value is between the low risk threshold and the medium risk threshold, it is determined to be a medium risk level. When the predicted value is between the medium risk threshold and the high risk threshold, it is determined to be a high risk level. When the predicted value exceeds the high risk threshold, it is determined to be a severe risk level.

Images