A method, system, device, and medium for predicting runoff pollutant concentrations

Abstract

The present application relates to the field of environmental science and technology, and discloses a runoff pollutant concentration prediction method, system, device and medium, the method comprising: for each to be predicted pollutant in the runoff sample, obtaining its environmental proxy data, constructing multi-dimensional prediction features based on rainfall data and environmental proxy data; constructing a composite feature under the mechanism of generation-accumulation-erosion; selecting an initial feature set from the multi-dimensional prediction features and the composite feature; jointly optimizing based on the initial feature set and the initial model parameter set to obtain a target feature set and a target model parameter set; training an initial prediction model based on the target feature set and the target model parameter set to obtain a target prediction model; and predicting based on the actual feature set using the target prediction model to obtain the predicted concentration. The present application has both the rationality of mechanism guidance and the accuracy of data-driven, can adapt to the scene of sparse runoff pollution data, improves the generalization ability of the model, and realizes efficient and accurate prediction of the pollutant concentration.

Description

Technical Field

[0001] This invention relates to the field of environmental science and technology, specifically to a method, system, equipment, and medium for predicting runoff pollutant concentrations. Background Technology

[0002] Urban stormwater runoff pollution is a key issue affecting urban water quality and restricting the utilization of stormwater resources. Accurately predicting runoff pollutant concentrations is crucial for implementing precise pollution control and risk warning.

[0003] Existing technologies mainly employ two types of models: one is a mechanistic model based on physical processes, such as the Storm Water Management Model (SWMM), which relies on physical laws to simulate the entire process of runoff pollution in order to achieve prediction; the other is a statistical and data-driven model based on historical data, which trains the model by collecting data on-site to predict pollutant concentrations during rainfall events.

[0004] However, mechanistic models are complex to construct, and key parameters such as accumulation and scour coefficients require backfit calibration using limited and difficult-to-obtain measured runoff data. Without sufficient calibration data, the model cannot be applied, and the parameter set is only suitable for specific regions and time periods, resulting in poor portability and generalization ability. Data-driven models are constrained by small sample sizes, and the randomness of rainfall events makes it difficult to obtain large-scale, high-quality training data. Directly applying machine learning models is prone to overfitting. In summary, existing technologies suffer from low prediction accuracy and poor generalization ability when predicting runoff pollutant concentrations. Summary of the Invention

[0005] This invention provides a method, system, device, and medium for predicting runoff pollutant concentrations, in order to solve the problems of low prediction accuracy and poor generalization ability in the existing technology when predicting runoff pollutant concentrations.

[0006] In a first aspect, the present invention provides a method for predicting runoff pollutant concentrations, the method comprising: For each pollutant to be predicted in the runoff sample, environmental proxy data of the pollutant to be predicted is obtained, and multidimensional prediction features are constructed based on the rainfall data and environmental proxy data of the runoff sample. Based on multidimensional prediction features, a composite feature of the pollutant to be predicted under the generation-accumulation-scouring mechanism is constructed. The initial feature set is selected from multidimensional predictive features and composite features; An initial prediction model for the pollutant to be predicted is constructed, and joint optimization is performed based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set. Based on the target feature set and the target model parameter set, an initial prediction model is trained to obtain the target prediction model; The actual feature set of the pollutant to be predicted is obtained, and the target prediction model is used to make predictions based on the actual feature set to obtain the predicted concentration of the pollutant.

[0007] This invention acquires environmental proxy data of the pollutants to be predicted and combines it with rainfall data from runoff samples to construct multidimensional predictive features, filling the information gap caused by random rainfall events. It constructs composite features of the pollutants to be predicted under the generation-accumulation-washout mechanism, systematically injecting prior knowledge from the field of environmental science into the data-driven model. This allows mechanistic analysis to guide data modeling, enabling the model to learn more accurately the intrinsic patterns that conform to the physical laws of runoff pollution under small sample conditions, and ensuring that the model's concentration predictions closely match the actual pollution migration and transformation process. An initial feature set with the highest signal-to-noise ratio for modeling is selected from the multidimensional and composite features, while avoiding overfitting due to too many feature dimensions. Joint optimization is performed based on the initial feature set and the initial model parameter set of the initial predictive model to obtain the target feature set and target model parameter set. This achieves the co-evolution of feature expression and model structure, thereby training the model to obtain the target predictive model. This model combines the rationality of mechanistic guidance with the accuracy of data-driven approaches, adapting to scenarios with sparse runoff pollution data, improving the model's generalization ability, and achieving efficient and accurate prediction of pollutant concentrations.

[0008] In one alternative implementation, the multidimensional predictive features include a base cumulative factor, rainfall dynamic variables, and surface runoff generation capacity variables. Based on multidimensional prediction features, a composite feature of the pollutant to be predicted under the generation-accumulation-fluedown mechanism is constructed, including: Identify the sources and inhibitory factors of the pollutants to be predicted. Inhibition factors refer to environmental factors that inhibit the surface accumulation of the pollutants to be predicted. Extract the first quantized feature corresponding to the source from the multidimensional prediction features, and extract the second quantized feature corresponding to the suppression factor; During the generation stage, a source strength expression is constructed based on the first quantization feature and its corresponding undetermined coefficients; In the accumulation phase, an effective accumulation potential expression is constructed based on the basic accumulation factor, the second quantitative feature and its corresponding undetermined coefficients; During the scour stage, a characteristic expression for scour is constructed based on rainfall dynamic variables and surface runoff generation capacity variables, as well as their respective undetermined coefficients. The source intensity expression, effective cumulative potential expression, and scour characteristic expression are used as composite features.

[0009] This embodiment transforms the mechanistic laws of the three key stages of pollutant generation, accumulation, and flushing into quantitative features that can be directly used by the model, realizing a systematic mapping from physical processes to mathematical expressions. This allows subsequent models to fully utilize prior knowledge in the domain, improve the interpretability and generalization ability of features, and provide high-quality feature inputs that are consistent in mechanism, logically complete, and dimensionally unified for accurate prediction of pollutant concentrations.

[0010] In one alternative implementation, an initial feature set is selected from multidimensional predictive features and composite features, including: Obtain the target concentration of the pollutant to be predicted; For each feature in the multidimensional prediction features, the correlation between the feature and the target concentration is evaluated to obtain a correlation score; Features with relevance scores below a preset threshold are removed, and the resulting multidimensional predicted features are used as the candidate feature set. An ensemble tree model is used to perform cross-validation on the candidate feature set, iteratively eliminating features that contribute the least to the model performance until the first preset stopping condition is met, thus obtaining the filtered candidate feature set. The selected candidate feature set and composite features are integrated into an initial feature set.

[0011] This embodiment uses correlation evaluation and cross-validation to screen features, providing comprehensive and non-redundant feature inputs for subsequent model training, which helps improve the model's generalization ability and computational efficiency.

[0012] In one optional implementation, joint optimization is performed based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set, including: Based on the value range of each undetermined coefficient in the initial feature set and the value range of each model parameter in the initial model parameter set, a hyperparameter space is constructed. Search for a set of parameter combinations in the hyperparameter space, including the value of each undetermined coefficient and the value of each model parameter; An initial feature set is instantiated based on parameter combination values, a temporary prediction model is configured based on parameter combination values, and a temporary prediction model is trained based on the initial feature set instantiated with the values. The temporary prediction model has the same model structure as the initial prediction model. Evaluate the performance of the temporary prediction model after training and obtain a performance score; The search direction is adjusted based on the performance score. Based on the adjusted search direction, the process returns to the step of searching for a set of parameter combinations in the hyperparameter space until the second preset stopping condition is met. The search is then stopped, and the parameter combination with the highest performance score is selected from all the searched parameter combinations to instantiate the initial feature set and the initial model parameter set, thus obtaining the target feature set and the target model parameter set.

[0013] This embodiment effectively avoids overfitting and improves the model's generalization ability by systematically searching and dynamically adjusting the direction based on the hyperparameter space, taking into account both global exploration and local fine-tuning optimization. It adapts to the prediction of pollutant concentration under different rainfall scenarios. Furthermore, through the joint optimization of the undetermined coefficients of features and model parameters, it achieves the co-evolution of feature representation and model structure.

[0014] In one optional implementation, obtaining the actual feature set of the pollutant to be predicted includes: For each feature in the target feature set, obtain real-time data of the pollutant to be predicted for that feature. Keeping the undetermined coefficients in the target feature set unchanged, an actual feature set is formed based on the real-time data of each feature.

[0015] This embodiment obtains the actual feature set based on the target feature set, ensuring the accuracy and stability of the prediction results.

[0016] In one optional implementation, after acquiring environmental proxy data of the pollutant to be predicted, the method further includes: Perform spatial and temporal alignment on the environment proxy data.

[0017] This embodiment helps to effectively integrate multi-source heterogeneous environmental proxy data by performing spatiotemporal alignment on the environmental proxy data, thus ensuring the accuracy of subsequent feature construction and model training.

[0018] In one alternative implementation, the method further includes: Using a target prediction model, the SHAP value of each feature in the actual feature set and the average SHAP value of all features are calculated. Interpretability analysis was performed based on the SHAP value and average SHAP value for each feature.

[0019] This embodiment uses SHAP values ​​for interpretability analysis, which not only improves the credibility of the prediction results, but also makes the originally complex black box model of machine learning transparent, so that its decision logic can be understood and verified by domain experts, providing intuitive decision support for risk warning and pollution control.

[0020] Secondly, the present invention provides a runoff pollutant concentration prediction system, the system comprising: The data fusion and quantification module is used to acquire environmental proxy data for each pollutant to be predicted in the runoff sample, and to construct multidimensional prediction features based on the rainfall data and environmental proxy data of the runoff sample. The domain feature engineering module is used to construct composite features of the pollutants to be predicted under the generation-accumulation-scouring mechanism based on multidimensional prediction features, and to select an initial feature set from the multidimensional prediction features and composite features; The joint optimization module is used to construct an initial prediction model for the pollutant to be predicted. It performs joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set. Based on the target feature set and the target model parameter set, the initial prediction model is trained to obtain the target prediction model. The prediction and interpretation module is used to obtain the actual feature set of the pollutant to be predicted, and to make predictions based on the actual feature set using the target prediction model to obtain the predicted concentration of the pollutant.

[0021] Thirdly, the present invention provides an electronic device, comprising: a memory and a processor, the memory and the processor being communicatively connected to each other, the memory storing computer instructions, and the processor executing the computer instructions to perform the runoff pollutant concentration prediction method of the first aspect or any corresponding embodiment described above.

[0022] Fourthly, the present invention provides a computer-readable storage medium storing computer instructions for causing a computer to execute the runoff pollutant concentration prediction method of the first aspect or any corresponding embodiment thereof. Attached Figure Description

[0023] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of a runoff pollutant concentration prediction system according to an embodiment of the present invention; Figure 2 This is a flowchart of a method for predicting runoff pollutant concentrations according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0025] 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.

[0026] It is understood that before using the technical solutions disclosed in the various embodiments of the present invention, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in the present invention and their authorization should be obtained in accordance with relevant laws and regulations through appropriate means.

[0027] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0028] Existing mechanistic models are complex to construct, and key parameters such as accumulation and scour coefficients require backfit calibration using limited and difficult-to-obtain measured runoff data. Without sufficient calibration data, the model cannot be applied, and the parameter sets are only suitable for specific regions and time periods, resulting in poor portability and generalization ability. Data-driven models are constrained by small sample sizes, and the randomness of rainfall events makes it difficult to obtain large-scale, high-quality training data. Directly applying machine learning models is prone to overfitting. In summary, existing technologies suffer from low prediction accuracy and poor generalization ability when predicting runoff pollutant concentrations.

[0029] This invention differs from the single path of pure mechanistic models or pure data-driven models by proposing a "mechanism-guided data-driven model" paradigm. With data-driven algorithms as the framework, it solves the problems of traditional mechanistic models in terms of parameter calibration difficulties and inability to describe complex processes in runoff scenarios. At the same time, by introducing composite features based on environmental science mechanisms, it overcomes the fundamental defect of pure data-driven models that are prone to overfitting due to the lack of prior knowledge guidance when dealing with small sample problems or complex pollutants with sparse data.

[0030] Furthermore, based on the "generation-accumulation-washout" physical framework commonly followed by pollutants when forming runoff pollution, this invention systematically generates composite features with high signal-to-noise ratio. It can effectively address the challenge that the environmental behavior of specific complex pollutants is diverse and difficult to characterize by a single original feature. By constructing composite features that can directly reflect their physical driving force, it fundamentally improves the model's learning ability under small sample conditions.

[0031] In addition, this invention alleviates information deficiency by fusing environmental proxy data and extracts the feature set most relevant to runoff pollution processes from multidimensional prediction features and composite features. This ensures that even when facing specific complex pollutants, the features input to the model have a higher signal-to-noise ratio than the original data, laying a good foundation for achieving better training results and model performance.

[0032] Figure 1 This is a schematic diagram of a runoff pollutant concentration prediction system according to an embodiment of the present invention. The system can be implemented based on general-purpose computer hardware (such as servers, personal computers, cloud platforms), such as... Figure 1 As shown, the data fusion and quantification module is used to acquire environmental proxy data for each pollutant to be predicted in the runoff sample, and construct multidimensional prediction features based on the rainfall data and environmental proxy data of the runoff sample; the domain feature engineering module is used to construct composite features of the pollutant to be predicted under the generation-accumulation-scour mechanism based on the multidimensional prediction features, and select an initial feature set from the multidimensional prediction features and composite features; the model joint optimization module is used to construct an initial prediction model for the pollutant to be predicted, and perform joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set, and train the initial prediction model based on the target feature set and the target model parameter set to obtain the target prediction model; the prediction and interpretation module is used to acquire the actual feature set of the pollutant to be predicted, and use the target prediction model to make predictions based on the actual feature set to obtain the predicted concentration of the pollutant to be predicted.

[0033] Specifically, the data fusion and quantification module integrates APIs (Application Programming Interfaces) for various data sources and spatial data processing tools (such as the GDAL and GEOS libraries), and supports manual data uploads by users. By acquiring and fusing rainfall data from runoff samples and environmental proxy data of the pollutants to be predicted, and then quantifying them, the original multidimensional prediction features are obtained. Next, the domain feature engineering module, based on the core physical process framework of pollutant "generation-accumulation-flue," uses its built-in composite feature calculation formula library to transform prior knowledge in the environmental science field into computable mathematical expressions, constructing composite features with clear physical meaning. These features are then combined with the multidimensional prediction features for feature selection, resulting in an initial feature set with a high signal-to-noise ratio. Finally, the model joint optimization module integrates automated hyperparameter tuning frameworks (such as Hyperopt and Optuna) and various machine learning model algorithm libraries (such as Scikit-learn, XGBoost, and LightGBM). It jointly optimizes the initial feature set and model hyperparameters to obtain the optimal target feature set and target model parameter set, which are then used to train the target prediction model. The prediction and interpretation module includes a prediction service interface that receives the actual feature set of the pollutants to be predicted from external input and inputs it into the target prediction model to complete the prediction of runoff pollutant concentrations. Through modular design, the complex data processing and modeling process is encapsulated into an automated, repeatable computer program, realizing an end-to-end solution from data input to prediction interpretation.

[0034] According to an embodiment of the present invention, a method for predicting runoff pollutant concentration is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0035] This embodiment provides a method for predicting runoff pollutant concentrations, which can be used in the aforementioned runoff pollutant concentration prediction system. Figure 2 This is a flowchart of a runoff pollutant concentration prediction method according to an embodiment of the present invention, such as... Figure 2 As shown, the process includes the following steps: Step S201: For each pollutant to be predicted in the runoff sample, obtain environmental proxy data for the pollutant to be predicted, and construct multidimensional prediction features based on the rainfall data and environmental proxy data of the runoff sample.

[0036] Specifically, due to the randomness of rainfall events, the basic data used for pollutant concentration prediction from runoff samples suffers from information sparsity. To fill this information gap, comprehensive and multidimensional environmental proxy data is matched to runoff samples to enrich feature dimensions and support subsequent model training, ensuring the effectiveness of concentration prediction. Simultaneously, different runoff pollutants exhibit significant differences in their physicochemical properties and pollution sources. Therefore, it is necessary to collect targeted proxy data that characterizes their entire migration and transformation process in the environment to achieve accurate concentration prediction of different pollutants in runoff samples.

[0037] For each pollutant to be predicted, environmental proxy data is selectively acquired and integrated through public data interfaces, remote sensing data platforms, and other means. This includes, but is not limited to, high-resolution meteorological data, land use data, human activity intensity proxy data, and potential pollution source characterization data. At the same time, core rainfall-related data such as rainfall duration, rainfall amount, and previous dry days from the runoff sample itself are integrated. All the integrated data are then quantified to form a unified, multi-dimensional prediction feature that includes meteorological, land use, human activity intensity, and traffic information. This provides a comprehensive and high-quality data foundation for the subsequent accurate prediction of runoff pollutant concentrations.

[0038] Taking benzene compounds as an example, the pollutants to be predicted have the physical and chemical characteristics of high volatility, moderate water solubility, and pollution sources mainly from transportation and industrial emissions. To address its high volatility, temperature and wind speed were collected from high-resolution meteorological data. Higher temperatures and wind speeds exacerbate the volatilization loss of benzene compounds from the surface or water bodies, making them key environmental factors affecting their effective concentration during accumulation and scouring. Considering its strong correlation with transportation and industrial sources, traffic data and POI (Point of Interest) data were collected from potential pollution source characterization data, and nighttime light index was collected from human activity intensity proxy data. Traffic patterns and road information can directly characterize the emission intensity of mobile pollution sources, while the location and density of specific POIs such as gas stations and chemical plants can characterize the pollution potential of stationary pollution sources. The nighttime light index, as a comprehensive quantitative indicator of human activity intensity, can indirectly reflect the overall pollutant emission level of the region. Regarding its accumulation and scouring behavior on the surface, the proportion of impervious surface area was collected from land use data. The proportion of impervious surface area directly determines the efficiency of runoff generation during rainfall and the ability of pollutants to be washed into water bodies, making it a key interface parameter affecting the transfer of benzene compounds from the surface to the aquatic phase.

[0039] High-resolution meteorological data is precisely integrated spatiotemporally based on the geographic coordinates of the sampling points of runoff samples and sampling times accurate to the minute. It is preferably obtained from the ERA5 reanalysis dataset of the European Centre for Medium-Range Weather Forecasts (ECMWF). More broadly, it can also originate from any meteorological element data source that can provide precise spatiotemporal matching with the sampling points and has a preset time resolution (such as hourly or daily). Meteorological elements include, but are not limited to, temperature, wind speed, air pressure, and humidity. Data sources also include reanalysis data from other global or regional numerical weather prediction models, measured data from surface meteorological stations, and remote sensing inversion data. In addition to collecting core features such as temperature and wind speed, high-resolution meteorological data can also be expanded to include features such as air pressure and relative humidity to enrich meteorological information.

[0040] The nighttime light index is obtained using monthly nighttime light data (vcmsl version, GeoTIFF format) from VIIRS (Visible Infrared Imaging Radiometer Suite), or remote sensing image data acquired by other satellites that can effectively characterize the intensity of nighttime surface light. Specifically, a geographic buffer zone with a preset radius is established for each sampling point. The buffer zone radius can be flexibly adjusted based on actual factors such as the urban fabric and pollution source diffusion characteristics of the study area where the runoff sample is located. For example, it can be reasonably selected within a range of 100 to 2000 meters. Using spatial masking technology, all pixel values ​​within this buffer zone are accurately extracted from the raster data, and their average value is calculated as the nighttime light index for that sampling point, thereby quantifying the intensity of human activity around the runoff sample.

[0041] Land use data is preferably obtained from the annual land cover dataset of China at a resolution of 30 meters, but other remote sensing data products that can clearly provide land cover types can also be used. Specifically, a buffer analysis method consistent with the nighttime light index is adopted. Within a buffer of a preset radius at each sampling point, the pixel proportion of various land use types such as cultivated land, forest, water bodies, and impervious surfaces is statistically analyzed, and the proportion of impervious surfaces is accurately extracted as a key feature characterizing runoff generation capacity.

[0042] Traffic data and POI data are obtained by calling the application programming interface of the electronic map service provider. Spatial analysis is carried out on the buffer of the preset radius of each sampling point: (1) Obtain and aggregate traffic situation data, and calculate traffic load quantitative indicators such as smooth flow rate, congestion rate, blockage rate, average traffic flow, and peak hour traffic flow; (2) Obtain road network vector information, and systematically count the distribution and length characteristics of various types of roads (highways, main roads or secondary roads, etc.), such as total length and total area; (3) Accurately retrieve and count the number and spatial distribution characteristics of specific categories of POIs that are highly related to the pollution sources of the target pollutants, such as gas stations, industrial parks, and parking lots.

[0043] Alternatively, in addition to the above-mentioned scheme of using historical publicly available data to obtain rainfall data and environmental proxy data to expand the runoff sample, the required data can also be collected on-site to obtain higher quality multidimensional prediction features, thereby constructing a prediction model with better performance.

[0044] Step S202: Based on multidimensional prediction features, construct composite features of the pollutants to be predicted under the generation-accumulation-flushing mechanism.

[0045] Specifically, to address the issues of pure data-driven models lacking physical interpretability and being prone to learning data noise rather than actual pollution patterns under small sample conditions, this invention, based on the general physical process framework of generation-accumulation-washout that pollutants universally follow in urban environments, systematically injects prior knowledge from the field of environmental science into the data-driven model by constructing composite features that can directly characterize the key physical driving forces of runoff pollution. This enables mechanistic analysis to guide data modeling, allowing the model to learn more accurately the inherent patterns that conform to the physical laws of runoff pollution under small sample conditions, and ensuring that the model's concentration predictions accurately reflect the actual pollution migration and transformation process.

[0046] Step S203: Select an initial feature set from multidimensional prediction features and composite features.

[0047] Specifically, all composite features constructed based on the mechanism are merged with the original multidimensional prediction features to form an expanded feature pool that covers the entire pollution process and contains multidimensional information. To eliminate redundant feature information, reduce the computational cost of subsequent model training, and improve model learning efficiency, this expanded feature pool is screened to obtain the initial feature set with the highest signal-to-noise ratio for modeling. This lays the foundation for subsequent joint optimization of the model and avoids overfitting due to too many feature dimensions, ensuring the model's generalization ability.

[0048] Step S204: Construct an initial prediction model for the pollutant to be predicted, and perform joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set.

[0049] Specifically, due to the different physicochemical and source characteristics of various runoff pollutants, an initial prediction model is constructed for each pollutant to be predicted. An ensemble tree model, such as random forest or XGBoost (limited gradient boosting tree), is preferred. Other machine learning models capable of handling complex nonlinear relationships, such as support vector machines and deep neural networks, can also be selected. Because the composite features of different pollutants to be predicted differ, and undetermined coefficients are introduced during the construction of composite features, and although the initial prediction model has a defined basic structure, its model parameters are not specifically designed and cannot directly adapt to the concentration prediction requirements of the pollutant, joint optimization based on the initial feature set and the initial model parameter set is necessary to obtain the target feature set and the target model parameter set. Unlike traditional mechanistic model parameter calibration processes that rely on large amounts of measured data, joint optimization aims to transform the problem of finding undetermined coefficients and model parameters into a data-driven hyperparameter optimization problem under physical constraints. This allows machine learning algorithms to learn, based on limited measured data, the undetermined coefficients and model parameters that best represent the real runoff pollution process in the current specific scenario, thereby improving the data signal-to-noise ratio. This enables the model to achieve high-precision prediction of pollutant concentrations even under data sparse conditions and significantly enhances the model's generalization ability.

[0050] Step S205: Based on the target feature set and the target model parameter set, train the initial prediction model to obtain the target prediction model.

[0051] Specifically, the initial prediction model is configured with parameters according to the target model parameter set. Using the target feature set as training input, the configured initial prediction model is trained, allowing the model to fully learn the nonlinear mapping between target features and runoff pollutant concentrations. A cross-validation training strategy is employed to evaluate performance, ultimately yielding the optimal target prediction model. This model combines the rationality of mechanism-guided approaches with the accuracy of data-driven approaches, enabling it to adapt to scenarios with sparse runoff pollution data and achieve efficient and accurate prediction of pollutant concentrations.

[0052] Alternatively, other training strategies can be used for model training. The model training process is based on existing technology and will not be described in detail here.

[0053] Step S206: Obtain the actual feature set of the pollutant to be predicted, and use the target prediction model to make a prediction based on the actual feature set to obtain the predicted concentration of the pollutant to be predicted.

[0054] Specifically, in practical applications, for a new rainfall event, the actual feature set of the pollutant to be predicted in the rainfall event is obtained, and the target prediction model is used to make a prediction based on the actual feature set, so as to output the predicted concentration of the pollutant to be predicted under the rainfall event.

[0055] This invention acquires environmental proxy data of the pollutants to be predicted and combines it with rainfall data from runoff samples to construct multidimensional predictive features, filling the information gap caused by random rainfall events. It constructs composite features of the pollutants to be predicted under the generation-accumulation-washout mechanism, systematically injecting prior knowledge from the field of environmental science into the data-driven model. This allows mechanistic analysis to guide data modeling, enabling the model to learn more accurately the intrinsic patterns that conform to the physical laws of runoff pollution under small sample conditions, and ensuring that the model's concentration predictions closely match the actual pollution migration and transformation process. An initial feature set with the highest signal-to-noise ratio for modeling is selected from the multidimensional and composite features, while avoiding overfitting due to too many feature dimensions. Joint optimization is performed based on the initial feature set and the initial model parameter set of the initial predictive model to obtain the target feature set and target model parameter set. This achieves the co-evolution of feature expression and model structure, thereby training the model to obtain the target predictive model. This model combines the rationality of mechanistic guidance with the accuracy of data-driven approaches, adapting to scenarios with sparse runoff pollution data, improving the model's generalization ability, and achieving efficient and accurate prediction of pollutant concentrations.

[0056] This embodiment provides a method for predicting runoff pollutant concentrations, which can be used in the aforementioned runoff pollutant concentration prediction system. The method specifically includes the following steps: Step S301: For each pollutant to be predicted in the runoff sample, acquire environmental proxy data for the pollutant, and construct multidimensional prediction features based on the rainfall data and environmental proxy data of the runoff sample. For details, please refer to [link to details]. Figure 2 Step S201 of the illustrated embodiment will not be described again here.

[0057] In some optional implementations, after obtaining the environmental proxy data of the pollutant to be predicted in step S301, the method further includes: Step S302: Spatial and temporal alignment of the environment agent data.

[0058] Specifically, since environmental proxy data comes from multiple sources, there are issues such as inconsistent geographical coordinate systems and inconsistent time granularity. In order to achieve effective integration of multi-source heterogeneous environmental proxy data and ensure the accuracy of subsequent feature construction and model training, it is necessary to perform spatial and temporal alignment on the environmental proxy data so that all data can be accurately matched with runoff samples in the spatiotemporal dimensions.

[0059] When performing spatial alignment, GIS (Geographic Information System) spatial analysis methods are used to first unify all raster and vector data containing geographic coordinates in the environmental proxy data to the same geographic coordinate system (such as WGS-84). For raster data from different sources, grid unification is performed according to the data type and the resampling algorithm. That is, bilinear interpolation algorithm is used for continuous variables and nearest neighbor method is used for categorical variables, and they are unified to the preset reference grid. For vector data such as POI and road network, a geographic buffer with a preset radius is established for the sampling points of runoff samples, and the geometric or quantitative characteristics of vector data within the buffer (such as total road length and number of POIs) are calculated to achieve spatial association matching between vector data and sample points.

[0060] When performing time alignment, targeted processing is applied based on the rainfall event to which the runoff sample belongs: For hourly or minute-by-minute meteorological data, which is high-frequency data dynamically correlated with the rainfall process, the start and end times of the rainfall event are used as time boundaries to extract the statistical characteristic values ​​of the meteorological data within this time period as feature data. The statistical characteristic values ​​can be selected as average, maximum, or cumulative values ​​according to data characteristics and forecasting needs. For monthly or annually updated coarse-grained data such as nighttime light and land use, which is low-frequency data without real-time dynamic correlation with the rainfall process, no time period statistics are required. The original values ​​of the time period to which the runoff sample sampling time belongs are directly extracted. That is, the nighttime light index of the month to which the sampling time falls is matched, and the land use data of the year to which the sampling time falls is matched. Through the above targeted time alignment processing, the time granularity differences of multi-source environmental proxy data are completely eliminated, and the standardized integration of various types of data in the time dimension is achieved. This ensures that each environmental proxy data feature can be accurately bound to the corresponding rainfall event and runoff sample sampling, ensuring the consistency and correlation of all data in the time dimension.

[0061] Optionally, after constructing multidimensional predictive features, preprocessing can be performed, including handling missing values ​​and standardizing numerical features.

[0062] Step S303: Based on multidimensional prediction features, construct composite features of the pollutants to be predicted under the generation-accumulation-scour mechanism. The multidimensional prediction features include basic accumulation factors, rainfall dynamic variables, and surface runoff generation capacity variables.

[0063] Specifically, step S303 includes: Step S3031: Determine the source and inhibitory factor of the pollutant to be predicted. The inhibitory factor represents the environmental factors that inhibit the surface accumulation of the pollutant to be predicted.

[0064] Specifically, different pollutants have different physicochemical and source characteristics, and these physicochemical characteristics determine the inhibitory factors. Therefore, it is necessary to determine the source and inhibitory factor for each pollutant to be predicted individually. The source refers to the main emission source type of the pollutant in the urban environment; the inhibitory factor refers to the key environmental factors that inhibit the accumulation of pollutants on the surface during the non-rainfalling period, such as volatilization and degradation. For example, heavy metals have the physicochemical characteristics of being non-volatile and highly adsorbent, and their main pollution sources are industrial and transportation sources. Land use type (such as the significant differences in the adsorption capacity of asphalt, concrete, and green space soil for heavy metals) and surface roughness are used as inhibitory factors. Benzene series compounds mainly originate from transportation sources and stationary sources such as industry and commerce, and have the physicochemical characteristics of high volatility. Wind speed and temperature are used as inhibitory factors. By accurately determining the pollution source and inhibitory factor for each pollutant to be predicted, a targeted mechanistic basis can be provided for the subsequent construction of composite features, making the constructed features more closely reflect the actual environmental behavior of the pollutants.

[0065] Step S3032: Extract the first quantization feature corresponding to the source from the multidimensional prediction features, and extract the second quantization feature corresponding to the suppression factor.

[0066] Specifically, abstract features are transformed into quantifiable features that the model can recognize. Specifically, based on the source of the pollutant to be predicted, the first quantifiable feature is extracted from the multidimensional prediction features. For example, when the source is a traffic source, variables characterizing road geometry and variables characterizing traffic activity can be extracted from the multidimensional prediction features. Specifically, the variable characterizing road geometry is the total length or total area of ​​each type of road within the buffer zone; the variable characterizing traffic activity is the average traffic flow or peak hour traffic flow of the roads within the buffer zone. When the source is a fixed source, the number of POIs of each type can be extracted from the multidimensional prediction features.

[0067] Based on the inhibitory factors of the pollutants to be predicted, the second quantitative features are extracted from the multidimensional prediction features. For example, when the inhibitory factors are temperature and wind speed, the specific values ​​of temperature and wind speed are extracted from the multidimensional prediction features; when the inhibitory factors are land use type and surface roughness, the corresponding quantitative representation values ​​are extracted from the multidimensional prediction features.

[0068] Step S3033: In the generation stage, a source strength expression is constructed based on the first quantization feature and its corresponding undetermined coefficients.

[0069] Specifically, during the pollutant generation stage, to quantify the potential emission intensity of pollutants to be predicted within a specific spatiotemporal range, the abstract pollution source intensity is transformed into a physically meaningful and calculable mathematical expression. This requires selecting an appropriate source intensity feature calculation formula based on the source of the pollutant to be predicted, substituting it into the first quantification feature, and introducing the undetermined coefficients corresponding to the first quantification feature to construct the corresponding source intensity expression. The undetermined coefficients serve as a crucial bridge connecting the physical laws of pollution with data-driven learning. They preserve the formula structure describing pollution source intensity as prior knowledge in the environmental field, while defining key intensity parameters that lack universal theoretical values, have high scenario specificity, and are difficult to measure directly as undetermined variables that can take values ​​within a reasonable range. The undetermined coefficients are mainly applied in three scenarios: quantification of process rates (such as volatilization and degradation rate constants), weighting the contribution intensity of features of different dimensions (such as the relative weights of different pollution sources), and nonlinear correction of environmental impacts (such as undetermined parameters in nonlinear functions).

[0070] Taking benzene series pollutants as an example, their sources are traffic sources and stationary sources. Equation (1) is the source intensity expression for traffic sources, and Equation (2) is the source intensity expression for stationary sources.

[0071] (1) In the formula, Indicates the first The variables representing road geometry corresponding to each road type; Indicates the first The road type weights corresponding to each road type; This represents the variables that characterize traffic activities.

[0072] (2) In the formula, Indicates the first The POI type weight corresponding to each POI type (such as gas station, chemical plant, auto repair shop, etc.) represents the correlation between the POI and the pollutant to be predicted; Indicates belonging to the first The number of POIs of each POI type.

[0073] The road type weight and POI type weight in the above formula are the undetermined coefficients that need to be optimized later. By constructing the source intensity expression, the potential emission intensity of pollutants at the generation stage is quantitatively and modeled, making the calculation of pollution source intensity fit the actual emission pattern of pollutants. At the same time, the introduction of undetermined coefficients provides space for subsequent optimization of feature weights based on measured data, ensuring the accuracy and scenario adaptability of the source intensity representation.

[0074] Step S3034: In the accumulation phase, an effective accumulation potential expression is constructed based on the basic accumulation factor, the second quantitative feature and its corresponding undetermined coefficients.

[0075] Specifically, in the pollutant accumulation stage, in order to characterize the potential of pollutants to accumulate on the surface during non-rainfall periods, it is necessary to select an appropriate effective accumulation potential characteristic calculation formula based on the pollutants to be predicted, substitute the second quantitative characteristic, introduce the undetermined coefficients corresponding to the second quantitative characteristic, and introduce environmental factors as correction terms to construct the corresponding effective accumulation potential expression.

[0076] Equation (3) is the general form of the effective cumulative potential expression: (3) In the formula, This represents the basic cumulative factor, such as the number of dry days in the preceding period; The inhibition function, such as the exponential decay function or other functions that can reflect the physical trend of the inhibition factor increasing and the effective cumulative amount decreasing, is used to characterize the inhibition effect of the inhibition factor. Its specific expression is set according to different pollutants. This indicates an inhibitory factor.

[0077] If the pollutant to be predicted is a polycyclic aromatic hydrocarbon (PAH), it is sensitive to photodegradation. The inhibition factors are light intensity and duration, and the inhibition function can be selected from functions related to "sunshine hours" or "total solar radiation". If the pollutant to be predicted is a benzene series compound, a simplified inhibition function based on linear approximation can be used to characterize the volatilization effect. The inhibition factors are temperature and wind speed. The following equation (4) is the expression for the effective cumulative potential of benzene series compounds: (4) In the formula, This indicates the temperature weight, used to quantify the degree to which temperature inhibits the accumulation of benzene compounds; Indicates temperature; This indicates the wind speed weight, used to quantify the degree to which wind speed contributes to the inhibition of benzene series compound accumulation; Indicates wind speed.

[0078] The temperature weight and wind speed weight in equation (4) above are the undetermined coefficients that need to be optimized later. By constructing an effective accumulation potential expression, the quantitative and mechanistic characterization of the potential of pollutant accumulation stage is realized. It not only retains the influence of basic accumulation conditions such as the number of dry days in the early stage, but also accurately quantifies the inhibitory effect of the inhibitory factor. At the same time, by introducing the undetermined coefficients, the scene adaptability of the feature and the space for subsequent optimization are taken into account, ensuring that the composite feature can truly reflect the surface accumulation law of pollutants.

[0079] Step S3035: In the scour stage, a scour characteristic expression is constructed based on the rainfall dynamic variables and the surface runoff generation capacity variables, as well as their respective undetermined coefficients.

[0080] Specifically, during the pollutant flushing stage, to quantify the flushing and transport capacity of rainfall events on accumulated pollutants on the surface, it is necessary to select an appropriate flushing characteristic calculation formula based on the pollutants to be predicted, substitute rainfall dynamic variables and surface runoff generation capacity variables, and introduce their undetermined coefficients to construct the corresponding flushing characteristic expression. Among these, rainfall dynamic variables can be average / peak rainfall intensity, total rainfall, etc.; surface runoff generation capacity variables can be the percentage of impervious surface area, catchment slope, etc.

[0081] Equation (5) is the general form of the scouring characteristic expression: (5) In the formula, It represents the function between rainfall dynamic variables and surface runoff generating capacity variables. It can be selected as a product, weighted summation or other nonlinear combination, and its specific expression can be set according to different pollutants. Represents rainfall dynamic variables; This represents a variable indicating the surface runoff generation capacity.

[0082] If the pollutant to be predicted is a benzene series compound, then the product is chosen as... The average rainfall intensity is used as the rainfall dynamic variable, and the proportion of impervious surface area is used as the surface runoff generation capacity variable. Since the function form has been determined, the undetermined coefficient can be set to 1 in this specific implementation. The product of the two can represent the scouring efficiency, which serves as the scouring characteristic expression.

[0083] When the pollutant to be predicted is a benzene series compound, the scour characteristic expression can be obtained through multiplication without introducing additional unknown undetermined coefficients. However, when calculating the scour characteristic expression for other pollutants to be predicted, if... Unknown quantities, including weights and nonlinear correction terms, are defined as undetermined coefficients to optimize the contribution of the two variables to the scour characteristics. By constructing a scour characteristic expression, a quantitative and mechanistic characterization of the core driving factors in the pollutant scour stage is achieved. This approach takes into account both the dual influence of rainfall dynamics and surface runoff generation capacity, and the design of undetermined coefficients reserves space for subsequent data optimization, ensuring that the scour characteristics can accurately reflect the actual scour and transport process of surface pollutants by rainfall runoff.

[0084] Step S3036: The source intensity expression, the effective cumulative potential expression, and the scour characteristic expression are used as composite features.

[0085] Specifically, the expressions obtained from the above three stages are used as composite features, which transforms the mechanistic laws of the three key stages of pollutant generation, accumulation and flushing into quantitative features that can be directly used by the model. This achieves a systematic mapping from physical processes to mathematical expressions, enabling subsequent models to make full use of prior knowledge in the domain, improve the interpretability and generalization ability of features, and provide high-quality feature inputs that are consistent in mechanism, logically complete and dimensionally unified for accurate prediction of pollutant concentration.

[0086] Step S304: Select an initial feature set from multidimensional prediction features and composite features.

[0087] Specifically, step S304 includes: Step S3041: Obtain the target concentration of the pollutant to be predicted.

[0088] Specifically, based on the sampling point and sampling time of the runoff sample, the actual measured concentration of the pollutant to be predicted at the sampling point and corresponding sampling time is obtained and defined as the target concentration.

[0089] Step S3042: For each feature in the multidimensional prediction features, perform a correlation assessment between the feature and the target concentration to obtain a correlation score.

[0090] Specifically, since most composite features contain undetermined coefficients, the focus is on multidimensional predictive features without unknown parameters, and correlation analysis is performed on each feature. Specifically, correlation analysis methods such as Pearson correlation analysis or mutual information are used to assess the correlation between the feature and the target concentration, obtaining a correlation score. The higher the score, the stronger the feature's ability to characterize pollutant concentration and the more significant its impact.

[0091] Step S3043: Remove features with a correlation score lower than a preset threshold, and use the removed multidimensional prediction features as a candidate feature set.

[0092] Specifically, features with correlation scores below a preset threshold are removed, and only core features that are significantly correlated with the target concentration and have a strong explanatory power for changes in pollutant concentration are retained. The remaining features after screening are used as the candidate feature set.

[0093] Step S3044: Using an ensemble tree model, cross-validation is performed on the candidate feature set, and features that contribute the least to the model performance are iteratively eliminated until the first preset stopping condition is met, thus obtaining the filtered candidate feature set.

[0094] Specifically, an ensemble tree model is adopted, combined with grouped K-fold cross-validation. Through an iterative backward elimination strategy or other screening methods such as Lasso regression, the features that contribute the least to the model performance are removed in each iteration until the first preset stopping condition is reached, such as when the model cross-validation performance no longer improves, and the selected candidate feature set is obtained.

[0095] Step S3045: Integrate the filtered candidate feature set and composite features into an initial feature set.

[0096] Specifically, the selected candidate feature set and composite features are integrated into an initial feature set, providing comprehensive and non-redundant feature input for subsequent model training, which helps improve the model's generalization ability and computational efficiency.

[0097] Step S305: Construct an initial prediction model for the pollutant to be predicted, and perform joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set.

[0098] Specifically, step S305 above performs joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set, including: Step S3051: Based on the value range of each undetermined coefficient in the initial feature set and the value range of each model parameter in the initial model parameter set, construct a hyperparameter space that covers the feasible domain of all parameters, providing a complete and compliant search domain for subsequent automated parameter search.

[0099] Specifically, a hyperparameter space is constructed based on the range of values ​​for each undetermined coefficient and the range of values ​​for each model parameter (such as learning rate, tree depth, regularization strength, etc.).

[0100] Step S3052: Search for a set of parameter combinations from the hyperparameter space. The parameter combinations include the values ​​of each undetermined coefficient and the values ​​of each model parameter.

[0101] Specifically, automated hyperparameter tuning techniques such as Bayesian optimization and genetic algorithms are employed. The optimizer performs an intelligent search within the hyperparameter space and outputs a set of parameter combinations that include the specific values ​​of all undetermined coefficients and all model parameters.

[0102] Step S3053: Instantiate an initial feature set based on the parameter combination values, configure a temporary prediction model based on the parameter combination, train the temporary prediction model based on the initial feature set after value instantiation, and the temporary prediction model has the same model structure as the initial prediction model.

[0103] Specifically, the specific values ​​of the searched undetermined coefficients are filled into the corresponding features, and the specific values ​​of the searched model parameters are configured accordingly in the temporary prediction model. The configured temporary prediction model is trained using the initial feature set instantiated with the values.

[0104] Step S3054: Evaluate the performance of the trained temporary prediction model and obtain a performance score.

[0105] Specifically, the performance of the provisional prediction model is evaluated using a cross-validation strategy (such as grouped K-fold cross-validation), and the performance score is fed back to the optimizer. This performance can be R² (R-squared, coefficient of determination) or RMSE (Root Mean Squared Error).

[0106] Step S3055: Adjust the search direction based on the performance score. Based on the adjusted search direction, return to the step of searching for a set of parameter combinations in the hyperparameter space until the second preset stopping condition is met, stop the search, select the parameter combination with the highest performance score from all the searched parameter combinations, instantiate the initial feature set and the initial model parameter set, and obtain the target feature set and the target model parameter set.

[0107] Specifically, the optimizer adjusts the search direction for the next round based on the performance score. Following the new search direction, it returns to step S3052 to continue the search, gradually approaching the globally high-performance parameter combination that optimizes cross-validation performance. When the second preset stopping condition is met, such as reaching a preset number of iterations or performance convergence, the search stops. The combination with the highest performance score among all searched parameter combinations is determined, and its value is used to instantiate the initial feature set and the initial model parameter set, resulting in the target feature set and the target model parameter set. Through the joint optimization of undetermined coefficients and model parameters, the co-evolution of feature representation and model structure is achieved.

[0108] Optionally, if the model's performance score does not meet the preset engineering application standard when the preset number of iterations is reached, a higher level of iterative optimization can be initiated: adjust or add the construction method of composite features, or introduce new data sources to expand multidimensional prediction features, and then re-execute the subsequent steps.

[0109] Step S306: Based on the target feature set and the target model parameter set, train the initial prediction model to obtain the target prediction model. For details, please refer to [link to relevant documentation]. Figure 2 Step S205 of the illustrated embodiment will not be described again here.

[0110] Step S307: Obtain the actual feature set of the pollutant to be predicted, and use the target prediction model to make a prediction based on the actual feature set to obtain the predicted concentration of the pollutant to be predicted.

[0111] Specifically, step S307 above, which obtains the actual feature set of the pollutant to be predicted, includes: Step S3071: Obtain real-time data of the pollutant to be predicted for each feature in the target feature set.

[0112] Specifically, the target feature set is the optimal feature set obtained through screening, which contains the key feature dimensions that contribute the most to the prediction and have the lowest redundancy. When a new rainfall event occurs, real-time data of the pollutants to be predicted in the corresponding dimensions are collected based on the feature dimensions in this target feature set.

[0113] Step S3072: Keep the undetermined coefficients in the target feature set unchanged, and form the actual feature set based on the real-time data of each feature.

[0114] Specifically, the target feature set has completed the parameter instantiation process, and the undetermined coefficients have been filled with optimal values ​​and remain fixed. Based on this, the real-time data of each feature is filled into the corresponding positions, and finally, a practical feature set applicable to the current rainfall event is constructed, ensuring the accuracy and stability of the prediction results.

[0115] Step S308: Using the target prediction model, calculate the SHAP value of each feature in the actual feature set and the average SHAP value of all features.

[0116] Specifically, the prediction and interpretation module also includes an interpretability analysis engine with built-in interpretation libraries such as SHAP, which can calculate the SHAP value of each feature in the actual feature set and the average SHAP value of all features.

[0117] Step S309: Perform interpretability analysis based on the SHAP value and average SHAP value of each feature.

[0118] Specifically, the SHAP value of each feature quantifies its impact on the model's prediction results, while the average SHAP value quantifies the average impact of each feature on the model's output. By ranking all features based on their SHAP values, the most critical factors driving changes in runoff pollutant concentrations can be identified. For example, it can be verified whether a particular feature in a composite feature set dominates the model's decision-making, thus providing macro-level scientific insights for pollution source control and process management.

[0119] Optionally, the prediction and interpretation module also includes a visualization interface capable of displaying SHAP force plots or waterfall plots. This clearly demonstrates how each input feature pushes the prediction result up or down from the baseline value to the final predicted value in a specific prediction, providing a micro-level and precise attribution explanation for understanding the causes of specific high-pollution or low-pollution events. Furthermore, this interface can also be presented as a graphical web UI, displaying users the predicted values ​​of pollutant concentrations, historical trends, and interpretability analysis results, providing intuitive decision support for risk warning and pollution control.

[0120] Interpretability analysis not only improves the credibility of prediction results, but also makes the originally complex black box model of machine learning transparent, so that its decision-making logic can be understood and verified by domain experts.

[0121] This invention acquires environmental proxy data of the pollutants to be predicted and combines it with rainfall data from runoff samples to construct multidimensional predictive features, filling the information gap caused by random rainfall events. It constructs composite features of the pollutants to be predicted under the generation-accumulation-washout mechanism, systematically injecting prior knowledge from the field of environmental science into the data-driven model. This allows mechanistic analysis to guide data modeling, enabling the model to learn more accurately the intrinsic patterns that conform to the physical laws of runoff pollution under small sample conditions, and ensuring that the model's concentration predictions closely match the actual pollution migration and transformation process. An initial feature set with the highest signal-to-noise ratio for modeling is selected from the multidimensional and composite features, while avoiding overfitting due to too many feature dimensions. Joint optimization is performed based on the initial feature set and the initial model parameter set of the initial predictive model to obtain the target feature set and target model parameter set. This achieves the co-evolution of feature expression and model structure, thereby training the model to obtain the target predictive model. This model combines the rationality of mechanistic guidance with the accuracy of data-driven approaches, adapting to scenarios with sparse runoff pollution data, improving the model's generalization ability, and achieving efficient and accurate prediction of pollutant concentrations.

[0122] Compared with traditional technologies, the embodiments of the present invention can achieve the following significant beneficial effects when applied to the prediction of pollutant concentrations in urban stormwater runoff, especially when addressing the dilemma of predictive modeling of complex pollutants with sparse data: (1) Significantly improves the prediction accuracy and generalization ability of complex pollutants: Traditional technologies struggle to establish or perform poorly when facing runoff pollutants with sparse data and complex mechanisms, such as benzene series compounds. This invention, through data fusion and mechanism-guided feature engineering, enables the model to learn the patterns of runoff pollution more effectively. Theoretically, on the same small sample dataset, using this invention to predict the concentration of complex pollutants significantly improves the coefficient of determination compared to the baseline model, and the root mean square error is also significantly reduced, demonstrating practical application value. Specifically: Compared to a purely data-driven baseline model, the introduction of mechanism-based composite features provides higher-quality input to the model, effectively guiding it to learn physical laws rather than data noise, exhibiting stronger generalization ability; compared to traditional mechanism models, the automatic learning and optimization of key parameters through data-driven methods avoids dependence on large amounts of calibration data, significantly lowering the application threshold and achieving high prediction accuracy in data-sparse regions. (2) Enhance the interpretability of prediction models for complex pollutants: The embodiments of the present invention not only improve the prediction accuracy, but also introduce physically meaningful composite features and interpretability tools, so that the prediction process for complex pollutants such as benzene series compounds is no longer a black box, but can be clearly understood which physical process (generation, accumulation or flushing) dominates the change of pollutant concentration. This not only enhances the confidence in the prediction results, but also provides scientific insights for source control and process management of specific pollutants. (3) Achieving universality in predicting different runoff pollutants: The core of this invention lies in providing a universal methodological framework. Since the "generation-accumulation-washout" mechanism is the common process by which most surface pollutants form runoff pollution, it can be easily transferred and applied to the prediction of various types of complex pollutants (such as heavy metals, PAHs, etc.), and has strong scalability. (4) Lowering the application threshold for runoff pollution model construction: By avoiding the cumbersome parameter calibration process of traditional mechanistic models, the threshold and cost of building high-precision runoff pollution prediction models are reduced. In particular, for specific pollutants with high analysis costs and difficult data acquisition, it is a practical technical path with high cost-effectiveness that can still establish effective prediction models under limited data conditions.

[0123] Figure 3 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention.

[0124] The following is a detailed reference. Figure 3The diagram illustrates a structural schematic suitable for implementing an electronic device according to embodiments of the present invention. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 301, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 302 or a program loaded from memory 308 into random access memory (RAM) 303. The RAM 303 also stores various programs and data required for the operation of the electronic device. The processor 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0125] Typically, the following devices can be connected to I / O interface 305: input devices 306 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 307 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 308 including, for example, magnetic tapes, hard disks, etc.; and communication devices 309. Communication device 309 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 3 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0126] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device 309, or installed from a memory 308, or installed from a ROM 302. When the computer program is executed by the processor 301, it performs the functions defined in the runoff pollutant concentration prediction method of the embodiments of the present invention.

[0127] Figure 3 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0128] This invention also provides a computer-readable storage medium. The methods described above according to embodiments of the invention can be implemented in hardware or firmware, or implemented as computer code that can be recorded on a storage medium, or implemented as computer code downloaded via a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the runoff pollutant concentration prediction method shown in the above embodiments is implemented.

[0129] A portion of this invention can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to the invention through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0130] Although embodiments of the invention have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of the invention, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A method for predicting runoff pollutant concentrations, characterized in that, The method includes: For each pollutant to be predicted in the runoff sample, environmental proxy data of the pollutant to be predicted is obtained. Multidimensional prediction features are constructed based on the rainfall data of the runoff sample and the environmental proxy data. The environmental proxy data includes high-resolution meteorological data, land use data, human activity intensity proxy data and potential pollution source characterization data. The multidimensional prediction features include basic cumulative factors, rainfall dynamic variables and surface runoff generation capacity variables. Based on the multidimensional prediction features, a composite feature of the pollutant to be predicted under the generation-accumulation-flushing mechanism is constructed. An initial feature set is selected from the multidimensional predictive features and the composite features; An initial prediction model for the pollutant to be predicted is constructed, and joint optimization is performed based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set; Based on the target feature set and the target model parameter set, the initial prediction model is trained to obtain the target prediction model; Obtain the actual feature set of the pollutant to be predicted, and use the target prediction model to make a prediction based on the actual feature set to obtain the predicted concentration of the pollutant to be predicted; The construction of composite features of the pollutant to be predicted under the generation-accumulation-washout mechanism based on the multidimensional prediction features includes: The sources and inhibitory factors of the pollutant to be predicted are determined, wherein the inhibitory factors represent environmental factors that inhibit the surface accumulation of the pollutant to be predicted. Extract the first quantized feature corresponding to the source from the multidimensional prediction features, and extract the second quantized feature corresponding to the inhibition factor; During the generation stage, a source strength expression is constructed based on the first quantization feature and its corresponding undetermined coefficients; During the accumulation phase, an effective accumulation potential expression is constructed based on the basic accumulation factor, the second quantification feature, and its corresponding undetermined coefficients. During the scouring phase, a scouring characteristic expression is constructed based on the rainfall dynamic variables, the surface runoff generation capacity variables, and their respective undetermined coefficients. The source strength expression, the effective cumulative potential expression, and the scour characteristic expression are taken as the composite feature.

2. The method according to claim 1, characterized in that, The step of selecting an initial feature set from the multidimensional prediction features and the composite features includes: Obtain the target concentration of the pollutant to be predicted; For each feature in the multidimensional prediction features, the correlation between the feature and the target concentration is evaluated to obtain a correlation score; Features with relevance scores below a preset threshold are removed, and the resulting multidimensional predicted features are used as the candidate feature set. An ensemble tree model is used to perform cross-validation on the candidate feature set, and the features that contribute the least to the model performance are iteratively eliminated until the first preset stopping condition is met, thus obtaining the filtered candidate feature set. The selected candidate feature set and the composite feature are integrated into the initial feature set.

3. The method according to claim 1, characterized in that, The joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain the target feature set and the target model parameter set includes: Based on the value range of each undetermined coefficient in the initial feature set and the value range of each model parameter in the initial model parameter set, a hyperparameter space is constructed. Search for a set of parameter combinations from the hyperparameter space, the parameter combinations including the value of each undetermined coefficient and the value of each model parameter; The initial feature set is instantiated based on the parameter combination values, a temporary prediction model is configured based on the parameter combination, and the temporary prediction model is trained based on the instantiated initial feature set. The temporary prediction model has the same model structure as the initial prediction model. Evaluate the performance of the temporary prediction model after training and obtain a performance score; Based on the performance score, the search direction is adjusted. Based on the adjusted search direction, the process returns to the step of searching for a set of parameter combinations from the hyperparameter space until the second preset stopping condition is met. The search is then stopped, and the parameter combination with the highest performance score is selected from all the searched parameter combinations to instantiate the initial feature set and the initial model parameter set, thereby obtaining the target feature set and the target model parameter set.

4. The method according to claim 3, characterized in that, The step of obtaining the actual feature set of the pollutant to be predicted includes: According to each feature in the target feature set, obtain the real-time data of the pollutant to be predicted for that feature; Keeping the undetermined coefficients in the target feature set unchanged, the actual feature set is formed based on the real-time data of each feature.

5. The method according to claim 1, characterized in that, After acquiring the environmental proxy data of the pollutant to be predicted, the method further includes: The environmental proxy data is spatially and temporally aligned.

6. The method according to claim 1, characterized in that, The method further includes: Using the target prediction model, calculate the SHAP value of each feature in the actual feature set and the average SHAP value of all features; Interpretability analysis is performed based on the SHAP value of each feature and the average SHAP value.

7. A runoff pollutant concentration prediction system, characterized in that, The system includes: The data fusion and quantification module is used to acquire environmental proxy data for each pollutant to be predicted in the runoff sample, and to construct multidimensional prediction features based on the rainfall data of the runoff sample and the environmental proxy data. The environmental proxy data includes high-resolution meteorological data, land use data, human activity intensity proxy data and potential pollution source characterization data. The multidimensional prediction features include basic cumulative factors, rainfall dynamic variables and surface runoff generation capacity variables. The domain feature engineering module is used to construct composite features of the pollutant to be predicted under the generation-accumulation-washout mechanism based on the multidimensional prediction features, and to select an initial feature set from the multidimensional prediction features and the composite features; The model joint optimization module is used to construct an initial prediction model for the pollutant to be predicted, perform joint optimization based on the initial feature set and the initial model parameter set of the initial prediction model to obtain a target feature set and a target model parameter set, and train the initial prediction model based on the target feature set and the target model parameter set to obtain a target prediction model. The prediction and interpretation module is used to obtain the actual feature set of the pollutant to be predicted, and to use the target prediction model to make a prediction based on the actual feature set to obtain the predicted concentration of the pollutant to be predicted. Specifically, the domain feature engineering module is used for: The sources and inhibitory factors of the pollutant to be predicted are determined, wherein the inhibitory factors represent environmental factors that inhibit the surface accumulation of the pollutant to be predicted. Extract the first quantized feature corresponding to the source from the multidimensional prediction features, and extract the second quantized feature corresponding to the inhibition factor; During the generation stage, a source strength expression is constructed based on the first quantization feature and its corresponding undetermined coefficients; During the accumulation phase, an effective accumulation potential expression is constructed based on the basic accumulation factor, the second quantification feature, and its corresponding undetermined coefficients. During the scouring phase, a scouring characteristic expression is constructed based on the rainfall dynamic variables, the surface runoff generation capacity variables, and their respective undetermined coefficients. The source strength expression, the effective cumulative potential expression, and the scour characteristic expression are taken as the composite feature.

8. An electronic device, characterized in that, include: A memory and a processor are communicatively connected, the memory storing computer instructions, and the processor executing the computer instructions to perform the runoff pollutant concentration prediction method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing the computer to execute the runoff pollutant concentration prediction method according to any one of claims 1 to 6.

Images