A rapid correction method and system for pollutant emission distribution based on multi-model fusion

Abstract

The invention discloses a method and system for rapidly correcting the distribution of pollutant emissions by multi-model fusion, wherein the WRF simulation results are corrected by the lightGBM model to obtain the WRF field and the meteorological field; the emission field and the SMOKE field are obtained by processing the SMOKE model; the simulation field simulated by CMAQ is corrected to obtain the corrected field, and the parameters of the lightGBM model at this time are recorded and saved as a lightGBM corrected model; the meteorological field and the emission field are input into the SA‑GNN model and the model results are corrected by the lightGBM corrected model to obtain the forecast field; the difference field between the corrected field and the forecast field is compared, and the emission field is output when the difference field is less than the threshold, and the emission field is adjusted when it is greater than the threshold until the latest emission field is output when the difference field is less than the threshold. Through multi-model fusion, the basic static inventory is quickly corrected, the accuracy is improved, and the use error is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of air quality monitoring, and in particular to a method and system for rapidly correcting pollutant emission distribution by fusion of multiple models. Background Art

[0002] The atmospheric pollutant emission inventory is an important basis for measuring the activity of pollution sources in a region, and is also a key tool for formulating environmental policies and implementing air pollution control measures.

[0003] Calculating the emissions of various types of air pollutants based on emission coefficients and activity levels is the mainstream technology for calculating emission inventories. However, due to limited localized emission coefficient data, my country's emission inventory is mainly based on the emission coefficient database of the United States. However, due to differences in urbanization, fuel properties, production processes, etc., borrowing foreign parameters leads to great uncertainty in the emission inventory. At the same time, in the calculation, the average emission coefficient is often used to calculate the emission inventory, which further expands the error. In the end, the emission inventory formed is not only seriously delayed in time, but also static, long-term scale, and coarse spatial granularity. This leads to large errors in actual use.

[0004] At present, the methods for revising emission inventories mainly include empirical model method, numerical simulation method, inversion method, statistical model method and machine learning method. The empirical model is simple but lacks dynamic adaptability, the numerical simulation method is highly accurate but complex in calculation, the inversion method relies on observational data and has limited resolution, the statistical model is fast but its linear assumption limits its applicability, and the machine learning method performs well in dealing with complex nonlinear relationships, but conventional machine learning methods rely on big data and lack interpretability. Each method has its own limitations in practical applications, and it is usually necessary to combine multiple methods to improve the accuracy and efficiency of the correction. Summary of the invention

[0005] The purpose of the present invention is to provide a method and system for rapid correction of pollutant emission distribution by multi-model fusion, which can rapidly correct the basic static inventory by combining multi-model fusion, joint graph neural network (SA-GNN) with transmission numerical model and machine learning model (lightGBM), improve accuracy and reduce usage errors. It can significantly improve the simulation accuracy of pollutant distribution field and ensure more accurate correction of emission inventory.

[0006] To achieve the above objectives, this application adopts the following scheme:

[0007] On the one hand, the present invention provides a method for rapidly correcting the pollutant emission distribution by multi-model fusion, which specifically comprises the following steps:

[0008] S1. Use the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, use the lightGBM model to correct the WRF simulation results to obtain the WRF field, and convert the WRF field into a meteorological field;

[0009] S2. Based on the MEIC inventory, the SMOKE model is used to process the emission inventory corresponding to the primary pollutants in the area to be corrected to obtain the SMOKE field, and the SMOKE field is formatted to obtain the emission field;

[0010] S3. Based on the WRF field and SMOKE field, CMAQ is used to simulate the pollution field distribution and obtain the CMAQ simulated field;

[0011] S4. According to the data of the primary pollutant monitoring station to be corrected, the CMAQ simulation field is corrected using the lightGBM model to obtain the corrected field, and the parameters of the lightGBM model at this time are recorded and saved as a lightGBM corrected model;

[0012] S5. Input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field;

[0013] S6, compare the difference field between the revised field and the forecast field to see if it is less than the set threshold value. If not, calculate the ratio of the concentration value of the primary pollutant in the difference field to that in the revised field, and use the ratio to correct the emission field to obtain a new emission field. Repeat step S5, update the SA-GNN model according to the meteorological field and the new emission field, and obtain an updated forecast field.

[0014] S7. Compare the difference between the revised field and the updated forecast field, and repeat step S6 until the difference is less than the set threshold.

[0015] In some specific implementations, the specific process of step S5 is:

[0016] S51, gridding the area to be corrected, constructing a graph neural network of the SA-GNN model, and using each grid point in the area to be corrected as a node of the graph;

[0017] S52, taking the input meteorological field and emission field as the characteristic vector of each node;

[0018] S53. According to the vector of each node, the nonlinear spatiotemporal relationship among the emission field, meteorological field and primary pollutant distribution is captured through the graph neural network to obtain the simulation field corresponding to the primary pollutant.

[0019] In some specific implementations, the process of constructing the graph structure of the graph neural network is as follows:

[0020] Define each grid point in the area to be corrected as a node of the graph;

[0021] Build edges between nodes, and the weight of the edges is determined according to the meteorological conditions between the nodes;

[0022] Based on the constructed graph structure, a spatial relationship matrix is ​​generated, which represents the mutual relationships and influences between nodes.

[0023] In some specific implementations, the specific process of step S53 is:

[0024] S531. Capturing the spatial distribution of primary pollutants through graph neural networks;

[0025] S532, SA-GNN model constructs several graph convolutional layers to aggregate and update node feature vectors and extract spatiotemporal dependencies between nodes;

[0026] S533. According to the spatial distribution relationship of the primary pollutant and the spatiotemporal dependency relationship between nodes, the concentration of the primary pollutant is simulated through a fully connected layer to obtain a simulation field corresponding to the primary pollutant.

[0027] In some specific embodiments, the SA-GNN model introduces a self-attention mechanism in the graph convolution layer operation, and the self-attention mechanism automatically assigns the influence of neighboring nodes according to the feature vector weight of each node.

[0028] In some specific implementations, the process of obtaining a new emission field in step S6 is:

[0029] Distribute the ratio of the corrected field to the forecast field to each grid in the area to be corrected, and find the coefficient of each grid; each grid is multiplied by the coefficient * emission field to obtain a new emission field.

[0030] In some specific implementations, the specific process of step S1 is:

[0031] S11. Obtain the initial and boundary conditions of the area to be corrected through global or regional reanalysis data, perform format conversion and regional resampling according to the initial and boundary conditions through the WRF pre-processing system, and generate meteorological input data suitable for the correction area;

[0032] S12, simulating the meteorological field by configuring various influencing parameters required for WRF simulation to obtain WRF simulation results;

[0033] S13, compare the WRF simulation results with the measured meteorological data to obtain the error, use lightGBM to correct the error of the WRF simulation results, and obtain the WRF field;

[0034] S14. Use Python to extract the meteorological field variables in the WRF field to form a meteorological data set that can be input into SA-GNN, and use the meteorological data set as the meteorological field.

[0035] In some specific implementations, the specific process of step S2 is:

[0036] S21. Obtain the MEIC emission inventory corresponding to the primary pollutants in the area to be corrected, format the data of the emission inventory according to the SMOKE requirements, and generate an emission source file that meets the SMOKE model input requirements;

[0037] S22. Input the emission source file into the SMOKE model, use the emission processing module of the SMOKE model to process the point source and surface source of the emission source, and generate a gridded emission data file suitable for CMAQ input. Use the gridded emission data file as the SMOKE field, and use python to format the gridded emission data file to obtain the emission field.

[0038] In some specific implementations, the steps of step S3 are:

[0039] S31. Configure various influencing parameters required for CMAQ simulation to simulate the pollution field;

[0040] S32. Input the WRF field and the SMOKE field into CMAQ to simulate the distribution of the pollution field in the area to be corrected and obtain the CMAQ simulated field.

[0041] In a second aspect, the present application provides a rapid correction system for pollutant emission distribution based on multi-model fusion, comprising:

[0042] The meteorological field simulation module uses the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, and uses the lightGBM model to correct the WRF simulation results to obtain the WRF field and meteorological field;

[0043] The emission field simulation module uses the SMOKE model to process the emission inventory corresponding to the primary pollutants in the area to be corrected, and obtains the emission field and SMOKE field;

[0044] The pollution field simulation module uses CMAQ to simulate the pollution field distribution based on the WRF field and SMOKE field to obtain the CMAQ simulation field;

[0045] The correction module is used to correct the CMAQ simulation field results using the lightGBM model according to the data of the primary pollutant monitoring station, obtain the corrected field, and record the parameters of the lightGBM model at this time, and save it as a lightGBM correction model;

[0046] The forecast field simulation module is used to input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field;

[0047] The emission field correction module is used to compare whether the difference between the revised field and the forecast field is less than the set threshold. If not, the ratio of the difference to the concentration value of the primary pollutant in the revised field is calculated. The emission field is corrected by this ratio to obtain a new emission field. The meteorological field and the new emission field are re-input into the SA-GNN model in the forecast field simulation module, and corrected using the lightGBM correction model to obtain an updated forecast field. The difference between the revised field and the updated forecast field is compared again. If it is still greater, the forecast field is continuously updated until the difference is less than the set threshold.

[0048] The inventive concept of the present invention is:

[0049] The existing methods for revising emission inventories mainly include empirical model method, numerical simulation method, inversion method, statistical model method and machine learning method. However, the empirical model is simple but lacks dynamic adaptability, the numerical simulation method is highly accurate but complex in calculation, the inversion method relies on observational data and has limited resolution, the statistical model is fast but its linear assumption limits its applicability, and the machine learning method performs well in processing complex nonlinear relationships, but conventional machine learning methods rely on big data and lack interpretability.

[0050] Therefore, this application uses a method that combines graph neural network (SA-GNN) with transmission numerical model and machine learning model (lightGBM) to quickly correct the basic static inventory. On the one hand, in order to reduce the error of the numerical model, the lightGBM model is used to pre-correct the WRF simulation results and CMAQ simulation results. At this time, the meteorological result field and the primary pollutant distribution simulation field obtained can be considered accurate, but the execution time of the entire system will be very long, and it is impossible to correct the static inventory.

[0051] Therefore, in order to achieve rapid correction of the emission inventory, the present invention is based on the dual learning idea, and now simulates the spatial pollutant distribution based on WRF-SMOKE-CMAQ. It is believed that the CMAQ simulation field is the result of the correct relationship. At the same time, lightGBM is added to the WRF and CMAQ simulation results for data correction to minimize the simulation numerical error, but this step cannot reversely modify the emission field data; accordingly, the meteorological field and emission field converted by WRF-SMOKE are input into SA-GNN, which can quickly realize the simulation of the distribution field of pollutants, and at the same time, the distribution field is corrected by lightGBM, and then the difference between the SA-GNN corrected field and the CMAQ simulated field is compared to correct the emission field. In this way, only the SA-GNN simulation is updated, and there is no need to re-simulate WRF, SMOKE and CMAQ, so that rapid correction of the emission inventory can be achieved.

[0052] The present invention has the beneficial effects:

[0053] 1. Adaptive learning of complex spatiotemporal relationships

[0054] By taking advantage of the natural advantages of the SA-GNN model in dealing with complex spatiotemporal relationships, the nonlinear spatiotemporal relationships between emission sources, meteorological conditions and pollutant distribution responses are captured through graph neural networks, which can well reflect the actual pollutant distribution field and improve the accuracy and adaptability of the simulation.

[0055] 2. It can achieve rapid iterative optimization and quickly revise the emission inventory

[0056] By constructing a difference field, the emission field can be quickly corrected, and only the SA-GNN simulation is updated without repeating the WRF, SMOKE and CMAQ simulations. This can achieve rapid correction of the emission inventory and efficiently complete iterative optimization. Compared with traditional numerical simulation methods that require fine-tuning of the emission field in each iteration, updating only the SA-GNN simulation speeds up the correction process and makes the correction of the emission inventory faster.

[0057] 3. Capable of dynamic adjustment and high-resolution processing

[0058] This application can dynamically adjust the emission field at the grid level according to the pollutant diffusion characteristics in different time and space, and realize high-resolution emission inventory correction. Compared with the static emission inventory calculation method, this dynamic correction method can better cope with complex and changing environmental conditions.

[0059] 4. It can reduce the dependence on artificial assumptions and empirical parameters

[0060] Traditional emission inventory correction methods usually rely on a large number of empirical parameters and assumptions, while the SA-GNN-based correction method learns the laws of pollutant transmission and diffusion through data-driven models, reducing dependence on artificial assumptions and parameter settings, and improving the scientificity and accuracy of the results.

[0061] 5. Flexible scalability and robustness

[0062] This application can adaptively process different types of meteorological and emission field data and has good scalability. Whether it is applied to different regions or different types of pollutants, the method can be flexibly adjusted according to the input data and has strong generalization ability and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of a method for rapidly correcting pollutant emission distribution by fusion of multiple models provided in an embodiment of the present invention;

[0064] Figure 2 A logical schematic diagram of the integration of various models provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0065] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. The following description of at least one exemplary embodiment is actually only illustrative and is by no means intended to limit the present invention and its application or use. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0066] The relative arrangement of components and steps, the numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present invention unless specifically stated otherwise.

[0067] At the same time, it should be understood that for the convenience of description, the sizes of the various parts shown in the drawings are not drawn according to the actual proportional relationship.

[0068] Additionally, descriptions of well-known structures, functions, and configurations may be omitted for clarity and conciseness.One of ordinary skill in the art will recognize that various changes and modifications may be made to the examples described herein without departing from the spirit and scope of the present disclosure.

[0069] Technologies, methods, and apparatus known to ordinary technicians in the relevant field may not be discussed in detail, but where appropriate, such technologies, methods, and apparatus should be considered part of the authorization specification.

[0070] In all examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limiting. Therefore, other examples of the exemplary embodiments may have different values.

[0071] The shortcomings of existing emission inventory calculation methods are:

[0072] 1. Traditional methods based on emission factors and statistical data: Traditional emission inventory accounting methods usually rely on fixed emission factors. These factors are based on historical statistical data or laboratory experimental results to estimate the emissions of specific pollution sources. By combining emission factors with the activity level of a region or industry (such as fuel consumption, production, etc.), the emissions of the corresponding pollutants are calculated.

[0073] Advantages: Simple and easy, data can be obtained through statistics and historical records.

[0074] Disadvantages: Simple model assumptions: Traditional methods assume that emission factors do not change with time, location or activity conditions, ignoring the dynamics and complexity of actual emissions.

[0075] Insufficient accuracy: Since these methods cannot accurately capture the differences between different pollution sources, they will produce large errors when dealing with complex and dynamically changing pollution sources.

[0076] Reliance on large amounts of historical data: This makes it difficult to incorporate emerging pollution sources or rapidly changing conditions.

[0077] 2. Methods based on statistical models and numerical simulation: In order to improve the accuracy of accounting, some methods combine numerical simulation and statistical models to establish more complex emission inventory calculation methods. This type of method simulates the emission characteristics of different pollution sources by inputting emission data into mathematical models.

[0078] Advantages: Compared with the simple emission factor method, this type of method can handle more complex systems and take into account more factors that affect emissions.

[0079] Disadvantages: High computational complexity: These models require a lot of computing resources and it is difficult to get results quickly.

[0080] Incomplete data problem: The accuracy of these models is still limited for regions with insufficient data or emerging pollution sources.

[0081] Susceptible to uncertainty: Due to the complexity of the model, any parameter error will lead to inaccurate final accounting results.

[0082] In response to the above-mentioned defects, the current methods for correcting emission inventories mainly include empirical model methods, numerical simulation methods, inversion methods, statistical model methods, and machine learning methods. The empirical model is simple but lacks dynamic adaptability, the numerical simulation method is highly accurate but the calculation is complex, the inversion method relies on observational data and has limited resolution, the statistical model is fast but the linear assumption limits its applicability, and the machine learning method performs well in dealing with complex nonlinear relationships, but conventional machine learning methods rely on big data and lack interpretability. Each method has its own limitations in practical applications, and it is usually necessary to combine multiple methods to improve the accuracy and efficiency of correction. Therefore, this application proposes the following scheme:

[0083] Example 1

[0084] like Figure 1-Figure 2 As shown, this embodiment provides a method for quickly correcting the pollutant emission distribution by multi-model fusion, which specifically includes the following steps:

[0085] S1. Use the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, use the lightGBM model to correct the WRF simulation results to obtain the WRF field, and convert the WRF field into a meteorological field;

[0086] The specific process of step S1 is:

[0087] S11. Obtain the initial and boundary conditions of the area to be corrected through global or regional reanalysis data, perform format conversion and regional resampling according to the initial and boundary conditions through the WRF pre-processing system, and generate meteorological input data suitable for the correction area;

[0088] S12, simulating the meteorological field by configuring various influencing parameters required for WRF simulation to obtain WRF simulation results;

[0089] S13, compare the WRF simulation results with the measured meteorological data to obtain the error, use lightGBM to correct the error of the WRF simulation results, and obtain the WRF field;

[0090] S14. Use Python to extract the meteorological field variables in the WRF field to form a meteorological data set that can be input into SA-GNN, and use the meteorological data set as the meteorological field.

[0091] S2. Based on the MEIC inventory, the SMOKE model is used to process the emission inventory corresponding to the primary pollutants in the area to be corrected to obtain the SMOKE field, and the SMOKE field is formatted to obtain the emission field;

[0092] The specific process of step S2 is:

[0093] S21. Obtain the MEIC emission inventory corresponding to the primary pollutants in the area to be corrected, format the data of the emission inventory according to the SMOKE requirements, and generate an emission source file that meets the SMOKE model input requirements;

[0094] S22. Input the emission source file into the SMOKE model, use the emission processing module of the SMOKE model to process the point source and area source of the emission source, and generate a gridded emission data file suitable for CMAQ input. Use the gridded emission data file as the SMOKE field, and format the gridded emission data file to obtain the emission field.

[0095] S3. Based on the WRF field and SMOKE field, CMAQ is used to simulate the pollution field distribution and obtain the CMAQ simulated field;

[0096] The steps of step S3 are:

[0097] S31. Configure various influencing parameters required for CMAQ simulation to simulate the pollution field;

[0098] S32. Input the WRF field and the SMOKE field into CMAQ to simulate the distribution of the pollution field in the area to be corrected and obtain the CMAQ simulated field.

[0099] S4. According to the data of the primary pollutant monitoring station to be corrected, the CMAQ simulation field is corrected using the lightGBM model to obtain the corrected field, and the parameters of the lightGBM model at this time are recorded and saved as a lightGBM corrected model;

[0100] S5. Input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field;

[0101] The specific process of step S5 is:

[0102] S51, gridding the area to be corrected, constructing a graph neural network of the SA-GNN model, and using each grid point in the area to be corrected as a node of the graph;

[0103] The construction process of the graph structure of the graph neural network is as follows:

[0104] Define each grid point in the area to be corrected as a node of the graph;

[0105] Build edges between nodes, and the weight of the edges is determined according to the meteorological conditions between the nodes;

[0106] Based on the constructed graph structure, a spatial relationship matrix is ​​generated, which represents the mutual relationships and influences between nodes.

[0107] S52, taking the input meteorological field and emission field as the characteristic vector of each node;

[0108] S53. According to the vector of each node, the nonlinear spatiotemporal relationship among the emission field, meteorological field and primary pollutant distribution is captured through the graph neural network to obtain the simulation field corresponding to the primary pollutant.

[0109] The specific process of step S53 is:

[0110] S531. Capturing the spatial distribution of primary pollutants through graph neural networks;

[0111] S532. The SA-GNN model constructs several graph convolution layers to aggregate and update node feature vectors and extract the spatiotemporal dependencies between nodes. The SA-GNN model introduces a self-attention mechanism in the graph convolution layer operation. The self-attention mechanism automatically allocates the influence of neighboring nodes according to the feature vector weight of each node, thereby enhancing the model's ability to capture spatial dynamic changes.

[0112] S533. According to the spatial distribution relationship of the primary pollutant and the spatiotemporal dependency relationship between nodes, the concentration of the primary pollutant is simulated through a fully connected layer to obtain a simulation field corresponding to the primary pollutant.

[0113] S6, compare the difference field between the revised field and the forecast field to see if it is less than the set threshold value. If not, calculate the ratio of the concentration value of the primary pollutant in the difference field to that in the revised field, and use the ratio to correct the emission field to obtain a new emission field. Repeat step S5, update the SA-GNN model according to the meteorological field and the new emission field, and obtain an updated forecast field.

[0114] The process of obtaining a new emission site in step S6 is as follows:

[0115] Distribute the ratio of the corrected field to the forecast field to each grid in the area to be corrected, and find the coefficient of each grid; each grid is multiplied by the coefficient * emission field to obtain a new emission field.

[0116] S7. Compare the difference between the revised field and the updated forecast field, and repeat step S6 until the difference is less than the set threshold.

[0117] In order to better illustrate the implementation process of this embodiment, a specific process of revising the PM2.5 emission inventory of atmospheric primary pollutants is taken as an example. The specific process is as follows:

[0118] 1. WRF meteorological field simulation: Use WRF to simulate the meteorological field in the area to be corrected

[0119] (1) Data acquisition: Use global or regional reanalysis data (such as FNL / GDAS) to obtain initial and boundary conditions, and perform format conversion and regional resampling through the WRF preprocessing system (WPS) to generate meteorological input data suitable for the correction area.

[0120] (2) Parameterization scheme configuration: The WDM6 scheme is used for microphysical processes, the RRTM model is used for longwave radiation simulation, the Dudhia model is used for shortwave radiation processing, the KF scheme is used for cumulus parameterization, the MM5 Monin-Obukhov scheme is used for near-surface layer simulation, the Noah LSM model is used for land surface process simulation, the Boulac scheme is used for planetary boundary layer simulation, and the BEP model is used for urban canopy simulation.

[0121] (3) Result correction: The WRF simulation results were compared with the measured meteorological data, and differential correction was performed using lightGBM.

[0122] (4) Result output: Use Python to extract the corrected meteorological field variables, such as wind speed, wind direction, temperature, humidity, precipitation, air pressure, radiation, etc. to form the required data set (hereinafter referred to as the meteorological field); at the same time, convert the result format (hereinafter referred to as the WRF field) for use in CMAQ input.

[0123] 2. SMOKE emission field processing: Use SMOKE to process basic emission data

[0124] (1) Data preparation: Obtain and preprocess the MEIC emission inventory, format the emission data according to SMOKE requirements, mainly including basic data on space, time and species distribution, and generate emission source files that meet the input requirements.

[0125] Spatial allocation: Based on the geographic information of emission sources, the MEIC emission inventory is allocated to the grids of the area to be corrected in a gridded mode to ensure that the location of the emission source matches the area to be corrected.

[0126] Time allocation: Through the time allocation module, emission data with time resolutions such as years, months, and days are allocated to finer time steps to ensure consistency with the time scale of the air quality model.

[0127] Species allocation and conversion: Based on the chemical mechanism, the emission species in the MEIC emission inventory are converted into the species required by the atmospheric chemistry model (such as CMAQ), and the conversion is carried out through chemical species mapping.

[0128] (2) Emission processing and output: Use the SMOKE emission processing module to process point and area sources of emission sources and generate gridded emission data files suitable for CMAQ input (hereinafter referred to as SMOKE fields). At the same time, use Python to process the format and form the required data set (hereinafter referred to as emission fields)

[0129] 3. CMAQ pollution field simulation:

[0130] (1) Configuration of key parameterization schemes: The CB05 scheme is used for chemical mechanisms, the Tiedtke scheme is selected for convection processes, the MYJ scheme is used for boundary layer processes, the RRTMG scheme is used for radiation processes, the Aero6 scheme is selected for aerosol processes, and the NOAH model is used for surface processes.

[0131] (2) Based on the WRF field and SMOKE field, CMAQ is used to simulate the pollution field distribution, which is called the CMAQ simulated field;

[0132] 4. Pollution field correction: Using the PM2.5 concentration monitoring station data, the CMAQ simulation field results are corrected based on lightGBM. The obtained field is called the corrected field, and the lightGBM parameters are recorded as the lightGBM model (this model will be used for distribution simulation field correction later)

[0133] 5. Graph structure construction:

[0134] (1) Each grid point in the simulation area is defined as a node of the graph, and the node feature vector includes the meteorological field and emission field of the point.

[0135] (2) Edges are constructed between nodes. The weight of the edge is determined based on the physical distance between nodes, wind speed and direction, and other meteorological conditions. The edge weight between nodes in areas with higher wind speeds is higher.

[0136] (3) Based on the constructed graph structure, a spatial relationship matrix is ​​generated, which represents the relationships and influences between nodes.

[0137] 6. SA-GNN model construction steps

[0138] (1) Design the SA-GNN model architecture. The model input is the feature vector of each node (including meteorological field and emission field), and the spatial propagation relationship of PM2.5 is captured through the graph neural network.

[0139] (2) A self-attention mechanism is introduced in the graph convolution operation. The self-attention mechanism distributes the influence of neighboring nodes according to the feature weight of each node, thereby enhancing the model's ability to capture spatial dynamic changes.

[0140] (3) Aggregate and update node features through several graph convolution layers to extract the spatiotemporal dependencies between nodes. After convolution, simulate PM2.5 concentration through a fully connected layer.

[0141] (4) Using the lightGBM model (the algorithm model obtained in step 4), the PM2.5 simulation field of SA-GNN is corrected, and the resulting field is called the forecast field.

[0142] 7. Correction based on the difference between the corrected and forecasted fields

[0143] (1) Calculate the difference between the PM2.5 correction field and the forecast field, and compare the difference with the set threshold

[0144] (2) If the difference is greater than the set threshold, the ratio of the difference to the PM2.5 concentration in the corrected field is distributed to the grids. At this time, a certain coefficient is obtained on each grid.

[0145] (3) Each grid is passed through the coefficient * emission field to obtain new emission field data

[0146] (4) Then, using the original meteorological field + the new emission field and the trained SA-GNN, we can get a new PM2.5 forecast field

[0147] 8. Repeat 6 (3)-7 until the difference is less than the set threshold. The whole process is completed, and the new emission field obtained last time is the final emission data, completing the emission inventory correction.

[0148] Example 2

[0149] The embodiment of the present invention provides a multi-model fusion pollutant emission distribution rapid correction system, which applies the method in embodiment 1 and includes:

[0150] The meteorological field simulation module uses the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, and uses the lightGBM model to correct the WRF simulation results to obtain the WRF field and meteorological field;

[0151] The emission field simulation module uses the SMOKE model to process the emission inventory corresponding to the primary pollutants in the area to be corrected, and obtains the emission field and SMOKE field;

[0152] The pollution field simulation module uses CMAQ to simulate the pollution field distribution based on the WRF field and SMOKE field to obtain the CMAQ simulation field;

[0153] The correction module is used to correct the CMAQ simulation field results using the lightGBM model according to the data of the primary pollutant monitoring station, obtain the corrected field, and record the parameters of the lightGBM model at this time, and save it as a lightGBM correction model;

[0154] The forecast field simulation module is used to input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field;

[0155] The emission field correction module is used to compare whether the difference between the revised field and the forecast field is less than the set threshold. If not, the ratio of the difference to the concentration value of the primary pollutant in the revised field is calculated. The emission field is corrected by this ratio to obtain a new emission field. The meteorological field and the new emission field are re-input into the SA-GNN model in the forecast field simulation module, and corrected using the lightGBM correction model to obtain an updated forecast field. The difference between the revised field and the updated forecast field is compared again. If it is still greater, the forecast field is continuously updated until the difference is less than the set threshold.

[0156] The above description is only a preferred embodiment of the present invention and does not limit the present invention in any form. According to the technical essence of the present invention, within the spirit and principles of the present invention, any simple modification, equivalent replacement and improvement made to the above embodiment still falls within the protection scope of the technical solution of the present invention.

Claims

1. A rapid correction method for pollutant emission distribution based on multi-model fusion, characterized in that: The specific steps include: S1. Use the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, use the lightGBM model to correct the WRF simulation results to obtain the WRF field, and convert the WRF field into a meteorological field; S2. Based on the MEIC inventory, the SMOKE model is used to process the emission inventory corresponding to the primary pollutants in the area to be corrected to obtain the SMOKE field, and the SMOKE field is formatted to obtain the emission field; S3. Based on the WRF field and SMOKE field, CMAQ is used to simulate the pollution field distribution and obtain the CMAQ simulated field; S4. According to the data of the primary pollutant monitoring station to be corrected, the CMAQ simulation field is corrected using the lightGBM model to obtain the corrected field, and the parameters of the lightGBM model at this time are recorded and saved as a lightGBM corrected model; S5. Input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field; The specific process of step S5 is: S51, gridding the area to be corrected, constructing a graph neural network of the SA-GNN model, and using each grid point in the area to be corrected as a node of the graph; The construction process of the graph structure of the graph neural network is as follows: Define each grid point in the area to be corrected as a node of the graph; Build edges between nodes, and the weight of the edges is determined according to the meteorological conditions between the nodes; Based on the constructed graph structure, a spatial relationship matrix is ​​generated, which represents the mutual relationships and influences between nodes; S52, taking the input meteorological field and emission field as the characteristic vector of each node; S53, according to the vectors of each node, the nonlinear spatiotemporal relationship between the emission field, the meteorological field and the primary pollutant distribution is captured through the graph neural network to obtain the simulation field corresponding to the primary pollutant; The specific process of step S53 is: S531. Capturing the spatial distribution of primary pollutants through graph neural networks; S532, SA-GNN model constructs several graph convolutional layers to aggregate and update node feature vectors and extract spatiotemporal dependencies between nodes; S533, according to the spatial distribution relationship of the primary pollutant and the spatiotemporal dependency relationship between nodes, the concentration of the primary pollutant is simulated through a fully connected layer to obtain a simulation field corresponding to the primary pollutant; S6, compare the difference field between the revised field and the forecast field to see if it is less than the set threshold value. If not, calculate the ratio of the concentration value of the primary pollutant in the difference field to that in the revised field, and use the ratio to correct the emission field to obtain a new emission field. Repeat step S5, update the SA-GNN model according to the meteorological field and the new emission field, and obtain an updated forecast field. S7. Compare the difference between the revised field and the updated forecast field, and repeat step S6 until the difference is less than the set threshold.

2. According to the method of rapid correction of pollutant emission distribution by multi-model fusion according to claim 1, it is characterized in that: The SA-GNN model introduces a self-attention mechanism in the graph convolution layer operation, which automatically assigns the influence of neighboring nodes according to the feature vector weight of each node.

3. The rapid correction method for pollutant emission distribution based on multi-model fusion according to claim 1 is characterized in that: The process of obtaining a new emission site in step S6 is as follows: The ratio of the concentration values ​​of the primary pollutants in the difference field and the correction field is distributed to each grid in the area to be corrected, and the coefficient of each grid is calculated; each grid is passed through the coefficient * emission field to obtain a new emission field.

4. The method for rapid correction of pollutant emission distribution based on multi-model fusion according to claim 1 is characterized in that: The specific process of step S1 is: S11. Obtain the initial and boundary conditions of the area to be corrected through global or regional reanalysis data, perform format conversion and regional resampling according to the initial and boundary conditions through the WRF pre-processing system, and generate meteorological input data suitable for the correction area; S12, simulating the meteorological field by configuring various influencing parameters required for WRF simulation to obtain WRF simulation results; S13, compare the WRF simulation results with the measured meteorological data to obtain the error, use lightGBM to correct the error of the WRF simulation results, and obtain the WRF field; S14. Use Python to extract the meteorological field variables in the WRF field to form a meteorological data set that can be input into SA-GNN, and use the meteorological data set as the meteorological field.

5. The method for rapid correction of pollutant emission distribution based on multi-model fusion according to claim 1 is characterized in that: The specific process of step S2 is: S21. Obtain the MEIC emission inventory corresponding to the primary pollutants in the area to be corrected, format the data of the emission inventory according to the SMOKE requirements, and generate an emission source file that meets the SMOKE model input requirements; S22. Input the emission source file into the SMOKE model, use the emission processing module of the SMOKE model to process the point source and area source of the emission source, and generate a gridded emission data file suitable for CMAQ input. Use the gridded emission data file as the SMOKE field, and format the gridded emission data file to obtain the emission field.

6. The method for rapid correction of pollutant emission distribution based on multi-model fusion according to claim 1 is characterized in that: The steps of step S3 are: S31. Configure various influencing parameters required for CMAQ simulation to simulate the pollution field; S32. Input the WRF field and the SMOKE field into CMAQ to simulate the distribution of the pollution field in the area to be corrected and obtain the CMAQ simulated field.

7. A system for rapidly correcting the distribution of pollutant emissions by fusion of multiple models, using the method for rapidly correcting the distribution of pollutant emissions by fusion of multiple models as claimed in claim 1, characterized in that: include: The meteorological field simulation module uses the WRF model to simulate the weather in the area to be corrected to obtain the WRF simulation results, and uses the lightGBM model to correct the WRF simulation results to obtain the WRF field and meteorological field; The emission field simulation module uses the SMOKE model to process the emission inventory corresponding to the primary pollutants in the area to be corrected, and obtains the emission field and SMOKE field; The pollution field simulation module uses CMAQ to simulate the pollution field distribution based on the WRF field and SMOKE field to obtain the CMAQ simulation field; The correction module is used to correct the CMAQ simulation field results using the lightGBM model according to the data of the primary pollutant monitoring station, obtain the corrected field, and record the parameters of the lightGBM model at this time, and save it as a lightGBM correction model; The forecast field simulation module is used to input the meteorological field and emission field into the SA-GNN model to obtain the simulation field corresponding to the primary pollutant, and use the lightGBM correction model to correct the simulation field of the primary pollutant to obtain the forecast field; The emission field correction module is used to compare whether the difference between the revised field and the forecast field is less than the set threshold. If not, the ratio of the difference to the concentration value of the primary pollutant in the revised field is calculated. The emission field is corrected by this ratio to obtain a new emission field. The meteorological field and the new emission field are re-input into the SA-GNN model in the forecast field simulation module, and corrected using the lightGBM correction model to obtain an updated forecast field. The difference between the revised field and the updated forecast field is compared again. If it is still greater, the forecast field is continuously updated until the difference is less than the set threshold.

Images