A gridded prediction method and system for air pollution based on meteorological parameter optimization

Abstract

This invention discloses a gridded prediction method and system for atmospheric pollution based on meteorological parameter optimization. It constructs a mapping database of parameter combinations and meteorological results, matches recommended parameter combinations to regional meteorological forecast data within the mapping database, and outputs gridded meteorological forecast data corresponding to the recommended parameter combinations. Historical pollutant concentration monitoring data and historical meteorological data are used as observation instance sets to screen a preliminary candidate variable set. Causal verification is performed on the preliminary candidate variable set to obtain a single variable set, which is then functionally coupled to obtain a set of related variables. The single variable set and the set of related variables are input into a pollutant concentration prediction model, outputting the pollutant concentration prediction results. Based on the gridded meteorological forecast data, the pollutant concentration prediction results are interpolated into each grid to obtain the gridded distribution of pollutant concentrations. Based on mapping database matching, multi-dimensional pollutant feature screening, and dynamic mapping fusion methods, high-precision modeling of meteorological forecasts and pollution distribution is achieved.

Description

Technical Field

[0001] This invention relates to the field of atmospheric environment monitoring technology, specifically to a gridded prediction method and system for atmospheric pollution based on meteorological parameter optimization. Background Technology

[0002] The formation of air pollutants such as PM2.5 is influenced by many complex factors, and their spatial and temporal variations make it difficult for traditional single models to achieve sufficient predictive accuracy. In existing technologies, the selection of WRF parameterization schemes heavily relies on human experience, resulting in poor regional adaptability and low efficiency. Furthermore, PM2.5 prediction relies on traditional feature selection methods (such as Pearson correlation coefficient and analysis of variance), but these methods cannot distinguish spurious correlations between variables. Simultaneously, traditional machine learning models adjust weights through a single loss function (such as MSE), making it difficult to balance the conflicting objectives of online learning and ensemble optimization. Summary of the Invention

[0003] The purpose of this invention is to provide a gridded prediction method and system for air pollution based on meteorological parameter optimization. It proposes a WRF-based intelligent parameter matching, multi-dimensional pollutant feature screening, and dynamic mapping fusion method to achieve high-precision modeling of meteorological forecasts and pollution distribution, providing reliable technical support for the control of air pollution problems.

[0004] To achieve the above objectives, this application adopts the following approach:

[0005] On the one hand, this invention provides a gridded prediction method for air pollution based on meteorological parameter optimization, specifically including the following steps:

[0006] S1. Obtain historical pollutant concentration monitoring data, historical meteorological data, and regional meteorological forecast data from micro-monitoring stations within the target area;

[0007] S2. Construct a mapping database of parameter combinations and meteorological results based on the physical parameterization scheme of the WRF forecast model. Match the recommended parameter combinations corresponding to the regional meteorological forecast data in the mapping database, input the recommended parameter combinations into the WRF forecast model and run it to output gridded meteorological forecast data in the target area.

[0008] S3. Use historical pollutant concentration monitoring data and historical meteorological data as observation instance sets to screen a preliminary candidate variable set; perform causal verification on the preliminary candidate variable set to obtain a single variable set; and perform functional coupling on the single variable set to obtain a related variable set.

[0009] S4. Input the single variable set and the associated variable set into the pollutant concentration prediction model, and output the pollutant concentration prediction results;

[0010] S5. Map the micro-monitoring stations into the grid of the target area, and interpolate the pollutant concentration prediction results into each grid based on the gridded meteorological forecast data to obtain the grid distribution of pollutant concentration.

[0011] In some specific implementation schemes, the specific process of step S2 is as follows:

[0012] S21. Obtain GFS data within the target area, select multiple core parameter combinations from the physical parameterization scheme of the WRF forecast model, use GFS data as the initial input of the WRF forecast model for each core parameter combination, and output the corresponding meteorological forecast data.

[0013] S22. Store the core parameter combination and its corresponding meteorological forecast data in the mapping database in the form of parameter combination-meteorological result combination.

[0014] S23. Using a classification model, the core parameter combination is used as the classification label and the meteorological station's forecast data is used as the feature. In the mapping database, the parameter combination with the smallest matching error for the meteorological station's forecast data is used as the recommended parameter combination.

[0015] S24. Input the recommended parameter combination into the WRF forecast model and run it to output gridded weather forecast data.

[0016] In some specific implementation schemes, the specific process of obtaining a single variable set in step S3 is as follows:

[0017] S31. Using historical pollutant concentration monitoring data and historical meteorological data as observation instance sets, Bayesian neural network modeling is performed on the original variables. The original variables include geographical factors and meteorological factors. Through Monte Carlo sampling, the prediction variance contribution of each original variable is quantified. Based on the prediction variance contribution, candidate variables are selected from each original variable and stored in the preliminary candidate variable set.

[0018] S32. Perform causal verification on each variable in the preliminary candidate variable set to obtain a single variable set.

[0019] In some specific implementation schemes, the specific process in step S32 is as follows:

[0020] The mean intervention effect ATE of each variable in the initial candidate variable set is calculated using causal forest. The causal direction p of each variable is verified by counterfactual analysis. Variables with ATE > 0.1 and p < 0.05 are selected from the initial candidate variable set and saved as a single variable set.

[0021] In some specific implementation schemes, the process of obtaining the set of related variables is as follows:

[0022] S01. Use wrap-around feature selection to combine variables in a single variable set, and perform collaborative validation on the variable combinations to obtain multiple combinations of significantly associated variables.

[0023] S02. Divide the target area into regions, merge similar regions according to the variables in each region to obtain several sub-regions, use the sub-regions as input nodes of the ST-GCN model, and obtain the terrain's barrier coefficient for pollutant diffusion through the k-hop neighbor aggregation of the ST-GCN model.

[0024] S03. Use a conditional generative adversarial network to generate extreme weather scenarios, and use the combined effect of terrain on the barrier coefficient of pollutant diffusion to screen the combination of significant correlation variables to obtain a set of candidate variables;

[0025] S04. Filter out unreasonable variables in the candidate variable set through physical constraints to obtain the associated variable set.

[0026] In some specific implementation schemes, the specific process of step S01 is as follows:

[0027] The predictive gain of variable combinations is evaluated by model importance score. Based on the predictive gain, variable combinations with clear physical meaning are selected from multiple variable combinations to obtain multiple Top-K combined variables.

[0028] Calculate the conditional average intervention effect of each Top-K combination variable, then perform counterfactual analysis to simulate the joint intervention of the Top-K combination variables, calculate the synergistic effect between the conditional average intervention effect of each Top-K combination variable and the joint intervention, and output the Top-K combination variables that satisfy the synergistic effect condition as the combination of significant association variables.

[0029] In some specific implementation schemes, step S4 is performed as follows:

[0030] S41. Use FTRL to select high-contribution and high-stability variables from single variable sets and related variable sets to form a basic variable set;

[0031] S42. Using the basic variable set as input to the pollutant concentration prediction model, construct several weak classifiers in the pollutant concentration prediction model to quantify the predictive power of different variable combinations, and output the pollutant concentration prediction results.

[0032] In some specific implementation schemes, step S5 is performed as follows:

[0033] S51. Match the grid of the target area with the grid of the WRF forecast model, and map the latitude and longitude of the micro-monitoring stations into the grid of the WRF forecast model as the reference point for grid interpolation.

[0034] S52. Select the corresponding interpolation algorithm based on the meteorological and topographical factors of each grid.

[0035] S53. Based on the reference points, the pollutant concentration prediction results are inserted into each grid according to the interpolation algorithm.

[0036] In some specific implementations, the interpolation algorithm includes integrated kriging, inverse distance weighting, radial basis function and modified IDW algorithm. A dynamic selector is constructed to store the correspondence between different meteorological and topographic factors and the interpolation algorithm used. The topographic deviation in the topographic factors of the grid is corrected by a neural network. The corresponding interpolation algorithm is selected from the dynamic selector according to the meteorological factors and the corrected topographic factors.

[0037] Secondly, this application provides a gridded atmospheric pollution prediction system based on meteorological parameter optimization, comprising:

[0038] One or more processors;

[0039] A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement a gridded prediction method for atmospheric pollution based on meteorological parameters, as described in the first aspect.

[0040] The beneficial effects of this invention are as follows:

[0041] By constructing a parameter combination-meteorological result mapping database, dynamic optimization and intelligent adaptation of the physical parameterization scheme for numerical weather prediction (WRF) were achieved. Through a collaborative mechanism of single variable screening and functional coupling of related variables, multi-scale dynamic optimization of pollutant concentration prediction was realized, significantly improving the model's generalization ability in complex meteorological and geographical scenarios. At the same time, through causal-physical joint driving, dynamic modeling of spatial heterogeneity, incremental weight allocation, and the adoption of multi-algorithm integration and intelligent decision-making mechanisms, the accuracy, efficiency, and scenario adaptability of pollutant gridded prediction were significantly improved. Attached Figure Description

[0042] Figure 1 A flowchart of a gridded prediction method for air pollution based on meteorological parameter optimization is provided for an embodiment of the present invention.

[0043] Figure 2 This is a schematic diagram of the gridded prediction process for air pollution provided in an embodiment of the present invention. Detailed Implementation

[0044] 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, and not all embodiments. The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the present invention or its application or use. 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.

[0045] Unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps described in these embodiments do not limit the scope of the invention.

[0046] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0047] Furthermore, for clarity and brevity, descriptions of well-known structures, functions, and configurations may have been omitted. Those skilled in the art will recognize that various changes and modifications can be made to the examples described herein without departing from the spirit and scope of this disclosure.

[0048] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0049] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0050] Example 1

[0051] like Figures 1-2 As shown, this embodiment provides a gridded prediction method for air pollution based on meteorological parameter optimization, specifically including the following steps:

[0052] 1. Data Processing

[0053] S1. Obtain regional geographic data, pollutant emission data, historical pollutant concentration monitoring data from micro-monitoring stations, historical meteorological data, and regional weather forecast data within the target area, including:

[0054] Regional static geographic data: topography, vegetation cover, land use type, etc.;

[0055] Historical pollutant concentration monitoring data from micro-monitoring stations: PM2.5 (data from multiple micro-monitoring stations in the target area and hourly data from each station).

[0056] Historical meteorological data: temperature, humidity, air pressure, wind speed, wind direction, dew point temperature, and rainfall (data from micro-meteorological stations in the target area and hourly data from meteorological monitoring stations).

[0057] Regional weather forecast data from the meteorological data network: wind speed, rainfall, temperature, wind direction, and air pressure;

[0058] GFS meteorological data; input data required for WRF forecasting models;

[0059] The data obtained above is processed, including the following steps:

[0060] (1) Missing data: There are missing values ​​in the original dataset, which are filled by linear interpolation.

[0061] (2) Data anomalies: Outliers are treated as missing values ​​and are handled using data missing handling methods.

[0062] (3) Data normalization: Min-Max Normalization is used to normalize the data.

[0063] 2. Dynamic optimization of WRF parameters

[0064] S2. Construct a mapping database of parameter combinations and meteorological results based on the physical parameterization scheme of the WRF forecast model. Match the recommended parameter combinations corresponding to the regional meteorological forecast data in the mapping database, input the recommended parameter combinations into the WRF forecast model and run it to output gridded meteorological forecast data in the target area.

[0065] The specific process of step S2 is as follows:

[0066] S21. Obtain GFS data within the target area, select multiple core parameter combinations from the physical parameterization scheme of the WRF forecast model, use GFS data as the initial input of the WRF forecast model for each core parameter combination, and output the corresponding meteorological forecast data.

[0067] S22. Store the core parameter combination and its corresponding meteorological forecast data in the mapping database in the form of parameter combination-meteorological result combination.

[0068] S23. Using a classification model, the core parameter combination is used as the classification label and the meteorological station's forecast data is used as the feature. In the mapping database, the parameter combination with the smallest matching error for the meteorological station's forecast data is used as the recommended parameter combination.

[0069] S24. Input the recommended parameter combination into the WRF forecast model, and run it in conjunction with regional static geographic data to output gridded meteorological forecast data.

[0070] Among them, the physical parameterization schemes include, but are not limited to, the cumulus schemes: KF, GR, BMJ; and the planetary boundary layer schemes: MYJ, ACM2, Boulac, etc. When selecting multiple core parameter combinations, the contribution of parameters to the prediction error can be quantified by machine learning models (such as random forest models) to select core parameter combinations. The random forest model can be used to train the nonlinear relationship between parameter combinations and meteorological results. The input is the parameter scheme (such as cumulus scheme = KF, boundary layer scheme = MYJ), and the output is the meteorological prediction error (such as temperature error ΔT, precipitation error ΔP). Key parameters are selected according to feature importance (Gini index) (such as the cumulus scheme having the greatest impact on precipitation error (weight 0.35), the boundary layer scheme having a significant impact on wind speed error (weight 0.28); the KF cumulus scheme has a positive correlation with meteorological station precipitation data (r=0.72), and the MYJ boundary layer scheme has a negative correlation with temperature error (r=-0.65)). Using classification models such as XGBoost, parameter combinations are used as classification labels and meteorological forecast data are used as features to train the classification model. The parameter combinations with the smallest error under similar meteorological conditions in the historical database are then input into the WRF to output gridded meteorological forecast data.

[0071] 3. Pollutant concentration prediction (taking PM2.5 as an example)

[0072] S3. Screening of Single and Correlated Variables: Historical pollutant concentration monitoring data and historical meteorological data are used as observation instance sets to screen a preliminary candidate variable set; causal verification is performed on the preliminary candidate variables to obtain a set of single variables; functional coupling is performed on the single variables to generate a set of related variables; the specific process of step S3 is as follows:

[0073] Obtain a single set of variables:

[0074] S31. Data preparation and variable selection: Historical pollutant concentration monitoring data and historical meteorological data are used as observation instance sets. Bayesian neural network modeling is performed on the observation instance sets. Monte Carlo sampling is used to quantify the contribution of each variable to the prediction variance. Candidate variables are selected from each variable based on the contribution of the prediction variance and stored in the preliminary candidate variable set.

[0075] The collected and processed historical daily average pollutant concentration dataset and meteorological dataset were used as the observation instance set (e.g., temperature, wind speed, vegetation cover, etc.). Bayesian neural network modeling was performed on the observation instance set. Monte Carlo Dropout was used to sample 1000 times to quantify the contribution of each variable to the prediction variance (e.g., wind speed variance = 3.2%, vegetation cover variance = 8.5%). Then, by setting a threshold (e.g., variance < 5%), highly stable variables (wind speed is acceptable, vegetation cover is excluded) were selected as the preliminary candidate variable set.

[0076] S32. Causal effect verification: Perform causal verification on each variable in the preliminary candidate variable set to obtain a single variable set.

[0077] The specific process in step S32 is as follows:

[0078] The mean intervention effect ATE of each variable in the initial candidate variable set is calculated using causal forest. The causal direction p of each variable is verified by counterfactual analysis. Variables with ATE > 0.1 and p < 0.05 are selected from the initial candidate variable set and output as a single variable set.

[0079] For the variables that passed the stability verification in step S31 above, the average intervention effect (ATE) is calculated using causal forest. For example, an increase in wind speed of 1 m / s leads to a decrease in PM2.5 of 0.25 μg / m³. The causal direction is verified by counterfactual analysis (whether the PM2.5 concentration would significantly increase if the wind speed were reduced by 10% (p<0.05)). Variables with insignificant causal effects are removed (such as land use type, ATE=0.08, p=0.12). Variables with ATE>0.1 and p<0.05 are stored in a single variable set.

[0080] The process of obtaining the set of related variables is as follows:

[0081] S01. Use wrap-around feature selection to combine variables in a single variable set, and perform collaborative validation on the variable combinations to obtain multiple combinations of significantly associated variables.

[0082] The specific process of step S01 is as follows:

[0083] The predictive gain of variable combinations is evaluated by model importance score. Based on the predictive gain, variable combinations with clear physical meaning are selected from multiple variable combinations to obtain multiple Top-K combined variables.

[0084] Calculate the conditional average intervention effect of each Top-K combination variable, then perform counterfactual analysis to simulate the joint intervention of the Top-K combination variables, calculate the synergistic effect between the conditional average intervention effect of each Top-K combination variable and the joint intervention, and output the Top-K combination variables that satisfy the synergistic effect condition as the combination of significant association variables.

[0085] The variables selected based on single variables are selected using Wrapper Methods. The predictive gain of the variable combinations after pairwise or triadic combinations is evaluated by model importance scores (such as Gini importance of random forest and gain contribution of XGBoost). Combinations with clear physical meaning are prioritized (such as wind speed × inversion layer thickness, etc., not full permutation combinations, to avoid the curse of dimensionality), and the Top-K combined variables are output.

[0086] Conditional average intervention effect (CATE) is calculated for the Top-K combined variables, for example, to assess the sensitivity of PM2.5 concentration changes when wind speed increases by 1 m / s and inversion layer thickness is >100 m. Counterfactual analysis is then performed to simulate joint interventions of the variable combination (such as simultaneously reducing wind speed by 10% and closing an industrial area) to verify whether the synergistic effect is significant (p<0.05), and the set of significant associated variables is output.

[0087] S02. Spatial Relationship Modeling: Divide the target area into regions, merge similar regions according to the variables in each region to obtain sub-region grids, use the sub-region grids as input nodes of the ST-GCN model, and obtain the terrain's barrier coefficient for pollutant diffusion through the k-hop neighbor aggregation of the ST-GCN model.

[0088] Spatial heterogeneity is quantified using a GeoDetector (for preprocessing only). A heterogeneity index is output based on regional characteristics, dynamically adjusting the variance contribution threshold of previously selected variables. Sub-regions with similar characteristics (e.g., grids with vegetation cover > 60%) are merged to reduce the number of computational nodes while preserving internal heterogeneity. The merged sub-regions serve as input nodes for the ST-GCN. Through k-hop neighbor aggregation (satisfying the diffusion equation boundary conditions), the spatiotemporal coupling effect between variables is captured, and the terrain's hindering effect on pollutant diffusion is quantified. Weights are assigned to terrain hindering effects (e.g., +0.2 for mountain elevation > 500m).

[0089] S03. Extreme Scenario Generation: Extreme weather scenarios are generated using a conditional generative adversarial network. The combination of significant correlation variables is screened by combining the obstacle coefficient of terrain to pollutant diffusion to obtain a set of candidate variables.

[0090] Conditional Generative Adversarial Networks (cGANs) are used to generate extreme weather scenarios (such as stable weather combined with high industrial emissions). Unreasonable variables in the set of associated variables are filtered out through physical constraints (mass conservation equations). The generated data is then compared with the observations of historical heavy pollution events (such as during red alerts for smog) to perform a distribution similarity test (KL divergence < 0.1) to ensure the physical interpretability of the generated data, and finally, a set of candidate variables is obtained.

[0091] S04. Feature Optimization: The number of candidate variables is reduced by dynamic causal screening, thus lowering the computational complexity. Based on high-weight variables (with high ATE) in a single variable set, combined features (e.g., wind speed × terrain relief) are generated first, rather than full permutations, reducing feature dimensionality. The filtered candidate variable set is then used as the associated variable set.

[0092] Understandably, the above process achieves multi-scale dynamic optimization of PM2.5 concentration prediction through a synergistic mechanism of single variable screening and functional coupling of related variables, significantly improving the model's generalization ability in complex meteorological and geographical scenarios. At the same time, through the closed-loop technology of "causal driving variable screening → spatial heterogeneity modeling → dynamic weight allocation", it realizes the leap from "experience-driven" to "mechanism-data fusion" in PM2.5 concentration prediction.

[0093] S4. Input the single variable set and the associated variable set into the pollutant concentration prediction model, and output the pollutant concentration prediction results; the specific process of step S4 is as follows:

[0094] S41. Use FTRL to select high-contribution and high-stability variables from single variable sets and related variable sets to form a basic variable set;

[0095] S42. Using the basic variable set as input to the pollutant concentration prediction model, construct several weak classifiers in the pollutant concentration prediction model to quantify the predictive power of different variable combinations, and output the pollutant concentration prediction results.

[0096] Dynamic weight allocation

[0097] Based on an online learning framework, high-contribution and high-stability variables are selected using L1 regularization (sparseness) and L2 regularization (stability) of FTRL (Follow-the-Regularized-Leader) to form a secondary candidate set. The high-stability variables from the FTRL output are then input into the Adaboost model to construct weak classifiers (decision stumps). Each weak classifier focuses on different combinations of variables (e.g., wind speed × terrain undulation). Boosting iteratively optimizes the classifier weights: each training round generates a weak classifier based on the current sample weights (initially uniformly distributed); the classification error is calculated, and the classifier weights are refined—lower errors result in larger weights—thus dynamically strengthening key combinations; fixed sample weights (here, a sample refers to each independent observation point in the training dataset, each sample containing multiple variables and a target value) are used to achieve single-variable and related-variable fusion through weighted voting by the classifiers, thereby training the Adaboost model. New air quality data is input into a trained Adaboost model. The model combines highly stable variables selected by FTRL with classifier weights learned by Adaboost to dynamically calculate predicted values ​​(for example, in stable weather scenarios, reduced wind speed leads to a significant increase in the weight of the "wind speed × inversion layer thickness" combination). The final prediction results are then fused with the outputs of multiple classifiers through weighted voting. The optimization effect is evaluated using metrics such as mean squared error (MSE) and coefficient of determination (R²), forming a closed-loop feedback loop of "data input → weight update → prediction optimization".

[0098] 4. Dynamic fusion of gridded data

[0099] S5. Map the micro-monitoring stations into the grid of the target area, and interpolate the pollutant concentration prediction results into each grid based on the gridded meteorological forecast data to obtain the grid distribution of pollutant concentration.

[0100] The specific process of step S5 is as follows:

[0101] S51. Match the grid of the target area with the grid of the WRF forecast model, and map the latitude and longitude of the micro-monitoring stations into the grid of the WRF forecast model as the reference point for grid interpolation.

[0102] The latitude and longitude of PM2.5 prediction stations (i.e., micro-monitoring stations) are mapped to the WRF grid as the reference point for grid interpolation;

[0103] S52. Select the corresponding interpolation algorithm based on the meteorological and topographical factors of each grid.

[0104] It integrates multiple interpolation algorithms such as Kriging, inverse distance weighting, and radial basis function (covering different meteorological and terrain conditions), and designs a dynamic selector to build a multi-interpolation algorithm library. Based on grid meteorological conditions and terrain features, it selects the interpolation algorithm online through a lightweight reinforcement learning model (e.g., when the wind direction changes suddenly and the terrain is undulating, it automatically switches to the modified IDW algorithm that considers obstacles; Kriging interpolation is used under calm and stable weather conditions).

[0105] S53. Based on the benchmark points, the pollutant concentration prediction results are interpolated into each grid using an interpolation algorithm. Using the parameter association rules mentioned in step 3 (parameter association rules include causal effects, spatial heterogeneity, sensitivity of combined features, and feedback mechanism of dynamic learning framework), the influence weights of single variable sets (key variables) and associated variable sets (variable combinations) in model prediction and data fusion are dynamically adjusted. Then, the terrain bias is corrected by a BP neural network (e.g., the concentration in valley areas is reduced by 12%), and finally, a 1km×1km high-resolution PM2.5 grid distribution map is output.

[0106] Model Update: When WRF meteorological grid data is updated (e.g., a new forecast every 6 hours) or new PM2.5 monitoring data is added, the following process is automatically triggered: ① Rematching station and grid data; ② Dynamically adjusting weight coefficients (e.g., updating weights after wind speed changes); ③ Running a hybrid interpolation algorithm to generate a new distribution map, using a lightweight neural network model (e.g., MobileNet) to ensure that the 1km×1km grid update is completed within 10 seconds.

[0107] Example 2

[0108] This embodiment provides a gridded atmospheric pollution prediction system based on meteorological parameter optimization, including:

[0109] One or more processors;

[0110] A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement a gridded prediction method for atmospheric pollution based on meteorological parameters, as described in the first aspect.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Based on the technical essence of the present invention, any simple modifications, equivalent substitutions, and improvements made to the above embodiments within the spirit and principles of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A gridded prediction method for air pollution based on meteorological parameter optimization, characterized in that, Specifically, the following steps are included: S1. Obtain historical pollutant concentration monitoring data, historical meteorological data, and regional meteorological forecast data from micro-monitoring stations within the target area; S2. Construct a mapping database of parameter combinations and meteorological results based on the physical parameterization scheme of the WRF forecast model. Match the recommended parameter combinations corresponding to the regional meteorological forecast data in the mapping database, input the recommended parameter combinations into the WRF forecast model and run it to output gridded meteorological forecast data in the target area. S3. Use historical pollutant concentration monitoring data and historical meteorological data as observation instance sets to screen a preliminary candidate variable set; perform causal verification on the preliminary candidate variable set to obtain a single variable set; and perform functional coupling on the single variable set to obtain a related variable set. The specific process of obtaining a single variable set in step S3 is as follows: S31. Using historical pollutant concentration monitoring data and historical meteorological data as observation instance sets, Bayesian neural network modeling is performed on the original variables. The original variables include geographical factors and meteorological factors. Through Monte Carlo sampling, the prediction variance contribution of each original variable is quantified. Based on the prediction variance contribution, candidate variables are selected from each original variable and stored in the preliminary candidate variable set. S32. Perform causal verification on each variable in the preliminary candidate variable set to obtain a single variable set; The specific process in step S32 is as follows: The mean intervention effect ATE of each variable in the initial candidate variable set is calculated using causal forest. The causal direction p of each variable is verified by counterfactual analysis. Variables with ATE > 0.1 and p < 0.05 are selected from the initial candidate variable set and saved as a single variable set. The process of obtaining the set of related variables is as follows: S01. Use wrap-around feature selection to combine variables in a single variable set, and perform collaborative validation on the variable combinations to obtain multiple combinations of significantly associated variables. S02. Divide the target area into regions, merge similar regions according to the variables in each region to obtain several sub-regions, use the sub-regions as input nodes of the ST-GCN model, and obtain the terrain's barrier coefficient for pollutant diffusion through the k-hop neighbor aggregation of the ST-GCN model. S03. Use a conditional generative adversarial network to generate extreme weather scenarios, and use the combined effect of terrain on the barrier coefficient of pollutant diffusion to screen the combination of significant correlation variables to obtain a set of candidate variables; S04. Filter out unreasonable variables in the candidate variable set through physical constraints to obtain the set of related variables; S4. Input the single variable set and the associated variable set into the pollutant concentration prediction model, and output the pollutant concentration prediction results; S5. Map the micro-monitoring stations into the grid of the target area, and interpolate the pollutant concentration prediction results into each grid based on the gridded meteorological forecast data to obtain the grid distribution of pollutant concentration.

2. The gridded prediction method for air pollution based on meteorological parameter optimization according to claim 1, characterized in that, The specific process of step S2 is as follows: S21. Obtain GFS data within the target area, select multiple core parameter combinations from the physical parameterization scheme of the WRF forecast model, use GFS data as the initial input of the WRF forecast model for each core parameter combination, and output the corresponding meteorological forecast data. S22. Store the core parameter combination and its corresponding meteorological forecast data in the mapping database in the form of parameter combination-meteorological result combination. S23. Using a classification model, the core parameter combination is used as the classification label and the meteorological station's forecast data is used as the feature. In the mapping database, the parameter combination with the smallest matching error for the meteorological station's forecast data is used as the recommended parameter combination. S24. Input the recommended parameter combination into the WRF forecast model and run it to output gridded weather forecast data.

3. The gridded prediction method for air pollution based on meteorological parameter optimization according to claim 1, characterized in that, The specific process of step S01 is as follows: The predictive gain of variable combinations is evaluated by model importance score. Based on the predictive gain, variable combinations with clear physical meaning are selected from multiple variable combinations to obtain multiple Top-K combined variables. Calculate the conditional average intervention effect of each Top-K combination variable, then perform counterfactual analysis to simulate the joint intervention of the Top-K combination variables, calculate the synergistic effect between the conditional average intervention effect of each Top-K combination variable and the joint intervention, and output the Top-K combination variables that satisfy the synergistic effect condition as the combination of significant association variables.

4. The gridded prediction method for air pollution based on meteorological parameter optimization according to claim 1, characterized in that, The specific process of step S4 is as follows: S41. Use FTRL to select high-contribution and high-stability variables from single variable sets and related variable sets to form a basic variable set; S42. Using the basic variable set as input to the pollutant concentration prediction model, construct several weak classifiers in the pollutant concentration prediction model to quantify the predictive power of different variable combinations, and output the pollutant concentration prediction results.

5. The gridded prediction method for air pollution based on meteorological parameter optimization according to claim 1, characterized in that, The specific process of step S5 is as follows: S51. Match the grid of the target area with the grid of the WRF forecast model, and map the latitude and longitude of the micro-monitoring stations into the grid of the WRF forecast model as the reference point for grid interpolation. S52. Select the corresponding interpolation algorithm based on the meteorological and topographical factors of each grid. S53. Based on the reference points, the pollutant concentration prediction results are inserted into each grid according to the interpolation algorithm.

6. The gridded prediction method for air pollution based on meteorological parameter optimization according to claim 5, characterized in that, The interpolation algorithm includes integrated kriging, inverse distance weighting, radial basis function and modified IDW algorithm. A dynamic selector is constructed to store the correspondence between different meteorological factors and terrain factors and the interpolation algorithm used. The terrain bias in the terrain factor of the grid is corrected by neural network. The corresponding interpolation algorithm is selected from the dynamic selector according to the meteorological factors and the corrected terrain factors.

7. A gridded air pollution prediction system based on meteorological parameter optimization, characterized in that, include: One or more processors; A storage unit is used to store one or more programs, which, when executed by one or more processors, enable the one or more processors to implement a gridded prediction method for atmospheric pollution based on meteorological parameters as described in any one of claims 1-6.

Images