Fast evaluation method for mesoscale atmospheric pollutant transport migration based on convolutional neural network

Abstract

The present application relates to a kind of based on convolutional neural network's middle-scale atmospheric pollutant transport migration fast evaluation method, solve the technical problems such as high requirement of prior art to computing resource, long computing time.The method includes:1, obtain middle-scale meteorological original data and static topographic data;2, generate key meteorological element spatial distribution characteristic data;3, obtain the source parameter of simulation area and carry out normalization processing;4, rely on middle-scale air quality model obtains pollutant three-dimensional spatial characteristic data;5, generate sample data set;6, to sample data set normalization processing obtains time series data set;7, according to time series data set extraction corresponding feature as input parameter;8, build convolutional neural network model and set loss function value and carry out model training;9, judge training result;10, extract trained convolutional neural network model, and design corresponding interface and generate fast evaluation model;11, using the model for evaluation.

Description

Technical Field

[0001] This invention relates to methods for simulating atmospheric pollution diffusion, specifically to a rapid assessment method for mesoscale atmospheric pollution diffusion and migration based on convolutional neural networks—long short-term memory. Background Technology

[0002] The study of pollutant diffusion and migration in the atmospheric environment is of great significance for research fields such as emergency prediction, chemical process analysis, air pollution control, and even infectious disease prevention. The transport of pollutants in the atmosphere is largely influenced by atmospheric turbulence, topographic elevation, and underlying surface type; therefore, it is a typical mesoscale problem strongly coupled with complex three-dimensional meteorological and topographic parameters. How to conduct mesoscale simulation and prediction of atmospheric pollutant diffusion and migration is crucial for emergency response decisions and the pre-analysis of physical processes in the aforementioned research fields.

[0003] Currently, researchers have developed typical mesoscale assessment models to simulate the transport characteristics of pollutants under complex source terms, meteorological conditions, and topographical features. Most existing assessment models employ traditional numerical simulation methods based on mathematical physics equations. These typically require the Euler method to discretize the entire atmospheric space into a grid, followed by corresponding numerical solutions to solve the discretized numerical problem across the entire simulation area. While this assessment model possesses strong capabilities for detailed three-dimensional description of physical parameters, the Euler model is computationally intensive, often requiring several hours of computation, and suffers from limitations in numerical convergence. Therefore, it does not meet the need for rapid assessment of atmospheric pollutant transport processes under sudden events.

[0004] For example, Ulas Im, Kostandinos Markakis, Alper Unal, et al. published "Study of winter PM episode in Istanbul using the high resolution WRF / CMAQ modeling system" (Atmospheric Environment, 44, 3085-3094) in 2010. This paper mainly simulated the concentration distribution characteristics of particulate matter in Istanbul during winter based on the WRF-CMAQ mesoscale model.

[0005] For example, V. Simsek, L. Pozzoli, A. Unal, T. Kindap, M. Karaca, et al. published "Simulation of 137Zhenhui Ma, Tengyue Ma, Baosheng Wang, et al. Meso-scale numerical analysis for transport and deposition behaviors of radioactive aerosols under severe nuclear accident. Progress in Nuclear Energy, 150, 104314. (2022), and Zhenhui Ma, Zhiming Li, Xiuhuan Tang, et al. Meso-scale numerical analysis for transport and deposition behaviors of radioactive aerosols under severe nuclear accident. Frontiers in Environmental Science, 12:1455273. (2024) mainly focus on the evaluation of harmful substance diffusion and migration of energy facilities under accident conditions.

[0006] The mesoscale methods of the above evaluations can achieve the evaluation of atmospheric pollutant transport and migration to a certain extent, however, they all involve three-dimensional grid discretization and solving the N-S equation based on the Euler method, which requires high computing resources and long computing time, and cannot quickly give the calculation results in a short time, thus cannot be used for rapid evaluation of harmful substance distribution and migration in emergency and other scenarios. SUMMARY

[0007] The purpose of the present application is to solve the technical problems of the existing evaluation methods, such as high requirement for computing resources and long computing time, and to provide a mesoscale atmospheric pollutant transport and migration rapid evaluation method based on convolutional neural network.

[0008] To achieve the above-mentioned purpose, the technical solution provided by the present application is:

[0009] The application discloses a fast evaluation method for mesoscale atmospheric pollutant transmission migration based on a convolutional neural network.

[0010] [1] Obtain mesoscale meteorological original data and static terrain data of a simulation area respectively; the mesoscale meteorological original data include original meteorological reanalysis data and assimilation data of the simulation area;

[0011] [2] Based on the mesoscale meteorological original data and the static terrain data obtained in step [1], continuous numerical simulation is performed by relying on a mesoscale advanced meteorology model to generate key meteorological element spatial distribution characteristic data of the simulation area;

[0012] [3] Obtain source item parameters of the simulation area, perform normalization processing on the source item parameters, and then set the emission condition of the source item as hourly emission and the emission amount as unit intensity;

[0013] [4] Combine the key meteorological element spatial distribution characteristic data generated in step [2] and the normalized source item parameters after the normalization processing in step [3], and perform mesoscale atmospheric diffusion numerical simulation by relying on a mesoscale air quality model, so as to obtain pollutant three-dimensional spatial characteristic data of the simulation area under the normalized source item hourly emission condition;

[0014] [5] The key meteorological element spatial distribution characteristic data generated in step [2] and the pollutant three-dimensional spatial characteristic data obtained in step [4] are subjected to data labeling and data alignment respectively according to different time steps, so as to generate a sample data set;

[0015] [6] The sample data set generated in step [5] is subjected to normalization processing, and then is organized according to time sequence, so as to obtain a time sequence data set;

[0016] [7] Time sequence features and spatial features in the time sequence data set are extracted respectively, and then are taken as input parameters, and the input parameters are divided into a training set and a test set according to time sequence;

[0017] [8] A convolutional neural network model is built, a loss function value is set, the training set in step [7] is input into the convolutional neural network model for model training, and then a training result is output;

[0018] [9] Whether the loss function value in the training result in step [8] is consistent with the set loss function value is judged, if not, the convolutional neural network model in step [8] is iteratively optimized combined with the test set, and then model training is continued; if yes, step

[10] is performed;

[0019]

[10] extracting the trained convolutional neural network model, and designing corresponding meteorological data interface, terrain data interface and output data interface to generate a mesoscale atmospheric pollutant transport and migration rapid evaluation model;

[0020]

[11] inputting the actual parameters of the simulation area into the mesoscale atmospheric pollutant transport and migration rapid evaluation model of step

[10] to complete the rapid evaluation of mesoscale atmospheric pollutant transport and migration based on the convolutional neural network.

[0021] Further, in step 【2】, the continuous numerical simulation refers to the horizontal grid configuration of the simulation area by adopting a three-layer grid nesting method, and the innermost layer grid is used as the final simulation area to carry out 1-3 years of mesoscale meteorological simulation.

[0022] The key meteorological elements include temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation and boundary layer height.

[0023] Further, in step 【2】, the horizontal grid resolution of the innermost layer grid is 1-3 km.

[0024] Further, in step 【5】, the sample data set includes a plurality of independent sample data groups, and each sample data group includes meteorological emission information and pollutant concentration information.

[0025] Further, in step 【6】, the organization according to the chronological order refers to the organization of the normalized sample data set according to the chronological order by using the sliding window method.

[0026] Further, in step 【6】, the time series data set includes a plurality of time series data groups with different starting times and simulation lengths.

[0027] Further, in step 【8】, the training refers to the multi-layer stacked long short-term memory processing of the training set, and then the three-dimensional convolution operation.

[0028] Further, in step 【8】, the loss function value is the sum of squares of absolute values of concentration differences between the training output results and the traditional simulation results of all grid points in the calculation domain.

[0029] Further, in step 【9】, the iterative optimization refers to taking the loss function value as the iteration index of model training, and adjusting the hyperparameters according to the iteration results.

[0030] Further, in step

[11] , it further includes the step of evaluating the evaluation model and the evaluation results using statistical indicators after the evaluation is completed.

[0031] Compared with the prior art, the present application has the following beneficial effects:

[0032] 1. The mesoscale atmospheric pollutant transport migration rapid evaluation method based on the convolutional neural network provided by the application is based on the mesoscale meteorological original data and static terrain data of a simulation area, relies on a mesoscale advanced meteorology model and a mesoscale air quality model to obtain pollutant three-dimensional space characteristic data of the simulation area, and then the data is taken as an input parameter to be trained through a convolutional neural network to obtain a mesoscale atmospheric pollutant transport migration rapid evaluation model, so that rapid evaluation is realized. The application can fully exert the long short-term memory architecture advantage of the convolutional neural network, fully capture the long-term dependence relationship in the time of the pollutant diffusion process, improve the prediction efficiency, and improve the prediction accuracy of the pollutant.

[0033] 2. The mesoscale atmospheric pollutant transport migration rapid evaluation method based on the convolutional neural network provided by the application can complete the rapid evaluation of the pollutant three-dimensional distribution trend and evolution law of a future few days of a sudden event in a short time (a few minutes) using a single machine without the aid of high-performance server calculation, so as to meet the actual needs of future rapid simulation evaluation of multiple accident scenes.

[0034] 3. The mesoscale atmospheric pollutant transport migration rapid evaluation method based on the convolutional neural network provided by the application has a coupling interface with a general GIS geographic information system and three-dimensional meteorological reanalysis data, and provides a convenient condition for the business operation of the method and the model.

[0035] 4. The mesoscale atmospheric pollutant transport migration rapid evaluation method based on the convolutional neural network provided by the application can realize the deployment of the model on a server or a cloud platform / quasi-cloud platform in the future, and therefore supports real-time or near-real-time multi-scene pollutant rapid rolling prediction. BRIEF DESCRIPTION OF DRAWINGS

[0036] Figure 1 is a flowchart of an embodiment of the application;

[0037] Figure 2 is a structure diagram of a convolutional neural network model in the embodiment of the application;

[0038] Figure 3 is a data flow structure diagram of the convolutional neural network model in the embodiment of the application;

[0039] Figure 4 is a preliminary prediction result comparison diagram of the embodiment of the application and a traditional numerical simulation method. DETAILED DESCRIPTION

[0040] In order to make the purposes, advantages and characteristics of the present application clearer, the present application will be further described in detail below in combination with the drawings and specific embodiments. Those skilled in the art should understand that these embodiments are only used to explain the technical principles of the present application, and are not intended to limit the protection scope of the present application.

[0041] With the rapid development of artificial intelligence technology, emerging artificial intelligence learning technologies such as convolutional neural networks and long short-term memories gradually interact with traditional theoretical research fields. Such technologies are mainly used to find the correlation between different data, rather than exploring the causal relationship through traditional numerical simulation methods, which makes it possible to solve mesoscale problems away from traditional mathematical and physical algorithms, and to some extent promotes the development of rapid evaluation technology. The present application combines the convolutional neural network-long short-term memory architecture with the traditional mesoscale meteorology-atmospheric transport model, generates a mesoscale atmospheric pollution diffusion and migration rapid evaluation model through the generation, processing, model optimization and iteration of sample data, and meets the needs of rapid evaluation of atmospheric transport and consequences of pollutants in fields including but not limited to atmospheric environment, chemical industry, health disease prevention, etc.

[0042] As shown in Figure 1 The present embodiment provides a mesoscale atmospheric pollutant transport and migration rapid evaluation method based on a convolutional neural network, comprising the following steps:

[0043] 【1】Obtain the mesoscale meteorological original data and static terrain data of the simulation area respectively. The mesoscale meteorological original data includes original meteorological reanalysis data and assimilation data of the simulation area.

[0044] In the present embodiment, the source of the reanalysis data can be meteorological reanalysis data published by the National Environmental Prediction Center (NCEP), the European Weather Forecasting Center (ECMWF) or the China Meteorological Administration (CMA). The assimilation data can use surface / sounding data or actual observation data published by official agencies.

[0045] 【2】Based on the mesoscale meteorological original data and static terrain data obtained in step 【1】, perform continuous numerical simulation relying on a mesoscale advanced meteorology model to generate key meteorological element spatial distribution characteristic data of the simulation area. The key meteorological elements include temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation and boundary layer height, and the simulation results of the key meteorological elements have a time resolution of every hour.

[0046] In the present embodiment, the continuous numerical simulation means that the horizontal grid of the simulation area is constructed in a three-layer grid nesting manner, and the innermost layer grid is used as the final simulation area to carry out mesoscale meteorological simulation for 1-3 years. The horizontal grid resolution of the innermost layer grid is generally 1-3 km.

[0047] The continuous simulation results of the three-dimensional meteorological field of the entire simulation area for 1-3 years provide meteorological driving for the subsequent mesoscale atmospheric diffusion simulation. The simulation objects of mesoscale atmospheric pollution diffusion include gases and aerosols. For gases, the three-dimensional spatial concentration data of pollutants are mainly simulated, and for aerosol pollutants, in addition to spatial concentration distribution data, surface deposition calculation data also need to be obtained.

[0048] For aerosol pollutants, according to the aerodynamic distribution characteristics of aerosols of different particle sizes, aerosols of different particle sizes can be classified and discussed: 10 μm is taken as the boundary to divide large-particle-size aerosols and small-particle-size aerosols into two groups. Generally, small-particle-size aerosols obey a logarithmic normal distribution, so they can be considered as a continuous spectrum; for large-particle-size aerosols, particle size groups can be divided, each particle size group has a typical average particle size, and a uniform particle size and related aerodynamic property parameters are used for description.

[0049] 【3】Obtain the source term parameters of the simulation area, normalize them, and then set the emission conditions of the source term as hourly emission and unit intensity.

[0050] In this embodiment, the source term parameters used in the mesoscale air quality model need to be normalized in source strength and time (it is considered that pollutants are only emitted for one hour according to unit quantity), and then according to the continuous numerical simulation results of 1-3 years in step 【2】, the normalized source term data is added in the mesoscale air quality model according to the corresponding time sequence every hour. The concentration and deposition simulation data of each normalized source term are output in the model to release the results of the next 7 days.

[0051] 【4】Combine the spatial distribution characteristic data of key meteorological elements generated in step 【2】 and the normalized source term parameters after step 【3】, rely on the mesoscale air quality model to perform mesoscale atmospheric diffusion numerical simulation, so as to obtain the three-dimensional spatial characteristic data of pollutants in the simulation area under the condition of hourly emission of normalized source term.

[0052] Among them, the mesoscale atmospheric diffusion numerical simulation can be performed by the simulation method disclosed in CN114547890A, or other methods that can achieve such results can be used for numerical simulation.

[0053] 【5】The spatial distribution characteristic data of key meteorological elements generated in step 【2】 and the three-dimensional spatial characteristic data of pollutants obtained in step 【4】 are respectively marked and aligned according to different time steps, so as to generate sample data sets that meet the convolutional neural network-long short-term memory architecture.

[0054] The sample dataset consists of multiple independent sample data groups, and each sample data group contains complete input information (i.e., meteorological data, normalized source data and meteorological emissions arranged hourly in time series) and output information (i.e., pollutant concentration and deposition information).

[0055] [6] Normalize the sample dataset generated in step [5], and then organize it in chronological order to obtain a time-series dataset, thereby improving the stability of the subsequent convolutional neural network-long short-term memory architecture.

[0056] In this embodiment, organizing according to chronological order means using the sliding window method to organize the normalized sample dataset in chronological order. That is, the time length of each sample data group remains unchanged, while the starting time changes hour by hour, generating multiple time-series data groups with different starting times but the same simulation duration, thereby obtaining a time-series dataset suitable for the long short-term memory architecture.

[0057] [7] Extract the temporal and spatial features from the time series dataset respectively, and use them as input parameters for the subsequent convolutional neural network-long short-term memory architecture. Divide the input parameters into training and test sets according to the time sequence to ensure that the established fast evaluation method has strong generalization ability for different cases and scenarios.

[0058] The input parameters include meteorological input characteristic parameters and emission characteristic input parameters. The meteorological input characteristic parameters mainly include wind speed, wind direction, precipitation and boundary layer height, while the emission characteristic input parameters mainly include the latitude and longitude location, altitude and surrounding topographic information of the emission source.

[0059] 【8】, such as Figure 2 As shown, a convolutional neural network model is built and a loss function value is set. Then, the training set described in step [7] is input into the convolutional neural network model for model training, and the training results are output.

[0060] The convolutional neural network model in this embodiment is a convolutional neural network-long short-term memory (LSTM) structure, which combines the characteristics of convolutional neural networks and LSM networks. It mainly includes the following components: the input X at the current time step. t The hidden state H of the previous time step t-1 and cell state C t-1 This unit extracts spatial features through convolutional operations and includes four main gating mechanisms: the forget gate F... t The input gate I determines how much information about the previous cell state to retain. t Controls the writing of current input information; candidate cell states generate new candidate states through the tanh function; output gate O tdetermines the output of the current cell state. The cell state C t is updated by combining the previous state and the current candidate state, while the hidden state H t is generated based on the updated cell state, according to an activation function, which is used to introduce a nonlinear characteristic, i.e.: whose output ranges from 0 to 1; whose output ranges from -1 to 1.

[0061] This design enables the convolutional neural network-long short-term memory unit structure to effectively capture the temporal data and spatial information.

[0062] The training described in this embodiment refers to the multi-layer stacked long short-term memory processing of the training set, and then the data is fed back to the output layer of the above architecture through three-dimensional convolution operation. The loss function value is set as the sum of squares of the absolute values of the concentration differences between the training output results and the traditional simulation results in all grid points in the calculation domain. The output training result is the output parameter of the sample data, including the three-dimensional concentration distribution and surface deposition distribution of the gas and aerosol pollutants at the future time.

[0063] Figure 3 is a data flow structure diagram based on convolutional neural network-long short-term memory. The input includes u-direction wind speed, v-direction wind speed, boundary layer height, and emission point information. Here, concn represents the predicted concentration value at the nth time, uvn represents the wind vector parameter at the nth time, and pbln represents the boundary layer height at the nth time. The first layer of the convolutional neural network-long short-term memory unit receives the above input data and processes it through the convolutional neural network-long short-term memory unit to capture complex temporal characteristics and spatial information. Subsequently, the subsequent several layers further refine the features to better focus on key temporal information. The output of the convolutional neural network-long short-term memory unit will be transmitted to a three-dimensional convolution layer, which is responsible for further processing the extracted features and converting them into the final output. The output is a single-channel result, representing the predicted concentration.

[0064] 【9】、determines whether the loss function value in the training result of step 【8】 is consistent with the set loss function value. If it is consistent, step

[10] is performed; if it is not consistent, the convolutional neural network model of step 【8】 is iteratively optimized in combination with the test set, and then the model training is continued, so that the comprehensive evaluation index including the loss function meets the actual needs of the evaluation. The iterative optimization described here refers to using the loss function value as the iterative index of model training, and adjusting the hyperparameters according to the iterative results. The specific hyperparameters that need to be adjusted include but are not limited to learning rate, batch size, long short-term memory layer number, and unit number, etc.

[0065]

[10] extract the trained convolutional neural network model, and design corresponding meteorological data interface, terrain data interface and output data interface, and then generate a mesoscale atmospheric pollutant transport migration rapid evaluation model.

[0066]

[11] input the actual parameters of the simulation area (including the source term parameters of the actual simulation case, geographic information data and meteorological reanalysis data, etc.) into the mesoscale atmospheric pollutant transport migration rapid evaluation model of step

[10] to complete the rapid evaluation of mesoscale atmospheric pollutant transport migration based on convolutional neural network.

[0067] Figure 4 is a comparison chart of the preliminary prediction results of the prediction method of the embodiment and the traditional numerical simulation method, wherein the left chart is the concentration diffusion simulation result of the traditional numerical simulation method, and the right chart is the concentration diffusion simulation result generated by the rapid evaluation method based on convolutional neural network-long short-term memory under the same condition according to the application. It can be seen from the chart that the concentration diffusion simulation result generated by the rapid evaluation method is basically consistent with the result of the traditional simulation method in terms of diffusion trend, diffusion range and concentration characteristics, thereby verifying the rationality of the method.

[0068] In addition, statistical indicators can be used to further evaluate the evaluation model and the evaluation result after the evaluation is completed, and the statistical indicators for evaluation include but are not limited to mean square error, root mean square error, mean absolute error and correlation coefficient.

[0069] The rapid evaluation method of mesoscale atmospheric pollutant transport migration based on convolutional neural network provided by the application can deeply consider the influence of atmospheric turbulent motion and vertical variation on pollutant transport, and deeply fuse the sudden source term, complex three-dimensional meteorological field and complex terrain / underlying surface conditions, thereby realizing the numerical simulation of mesoscale three-dimensional atmospheric diffusion under the influence of multiple factors. At the same time, the single machine can be used to complete the rapid evaluation of the three-dimensional distribution trend and evolution law of pollutants in the future few days under the premise of no need to rely on high-performance server calculation in a short time (a few minutes), thereby meeting the actual needs of future rapid simulation evaluation of multiple accident scenes, and the model can be deployed on a server or a cloud platform / quasi-cloud platform in the future, so that real-time or near-real-time rapid prediction of multiple scenes of pollutants is supported.

[0070] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the application, and not to limit them; although the application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement to part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the application.

Claims

1. A method for fast evaluation of mesoscale atmospheric pollutant transport migration based on convolutional neural network, characterized in that, The method comprises the following steps: 【1】. Obtain mesoscale meteorological original data and static terrain data of a simulation area respectively; The mesoscale meteorological original data comprises original meteorological reanalysis data and assimilation data of the simulation area; 【2】. Based on the mesoscale meteorological original data and the static terrain data, perform continuous numerical simulation by relying on a mesoscale meteorology model to generate spatial distribution characteristic data of key meteorological elements of the simulation area; 【3】. Obtain source term parameters of the simulation area, normalize the source term parameters, and then set the emission condition of the source term as hourly emission and the emission amount as unit intensity; 【4】. Combine the spatial distribution characteristic data of the key meteorological elements generated in step 【2】 and the normalized source term parameters, perform mesoscale atmospheric diffusion numerical simulation by relying on a mesoscale air quality model, and thus obtain three-dimensional spatial characteristic data of pollutants of the simulation area under the hourly emission condition of the normalized source term; 【5】. Mark and align the spatial distribution characteristic data of the key meteorological elements generated in step 【2】 and the three-dimensional spatial characteristic data of the pollutants according to different time steps, and thus generate a sample data set; 【6】. Normalize the sample data set, and then organize the sample data set according to time sequence to obtain a time series data set; 【7】. Extract time sequence features and spatial features in the time series data set as input parameters, and divide the input parameters into a training set and a test set according to time sequence; 【8】. Build a convolutional neural network model, set a loss function value, input the training set into the convolutional neural network model for model training, and then output a training result; 【9】. Determine whether the loss function value in the training result is consistent with the set loss function value, if not, iteratively optimize the convolutional neural network model combined with the test set, and then continue model training; if yes, execute step 【10】; 【10】. Extract the trained convolutional neural network model, design corresponding meteorological data interfaces, terrain data interfaces and output data interfaces, and generate a mesoscale atmospheric pollutant transport and migration rapid evaluation model; 【11】. Input actual parameters of the simulation area into the mesoscale atmospheric pollutant transport and migration rapid evaluation model, and thus complete the convolutional neural network-based mesoscale atmospheric pollutant transport and migration rapid evaluation.

2. The method according to claim 1, wherein in step 【2】, the continuous numerical simulation is performed by adopting a three-layer grid nesting method to configure horizontal grids of the simulation area, and the innermost layer of the grids is taken as the final simulation area to perform mesoscale meteorological simulation for 1-3 years; and the key meteorological elements comprise temperature, humidity, wind speed, wind direction, atmospheric pressure, precipitation and boundary layer height.

3. The method according to claim 2, wherein in step 【2】, the horizontal grid resolution of the innermost layer of the grids is 1-3 km. ​ ​ ​ 4. The method according to claim 1, wherein the step of 【5】 comprises: the sample data set comprises a plurality of sample data groups which are independent of each other, and each sample data group comprises meteorological emission information and pollutant concentration information.

5. The method according to claim 1, wherein the step of 【6】 comprises: the step of organizing the normalized sample data set in chronological order is performed by using a sliding window method.

6. The method according to claim 5, wherein the step of 【6】 comprises: the time series data set comprises a plurality of time series data groups with different starting times and simulation durations.

7. The method according to claim 1, wherein the step of 【8】 comprises: the training is performed by first performing multi-layer stacked long short-term memory processing on the training set, and then performing three-dimensional convolution operation.

8. The method according to claim 7, wherein the step of 【8】 comprises: the loss function value is the sum of squares of absolute values of differences between training output results and concentration results of traditional simulation results at all grid points in the calculation domain.

9. The method according to claim 1, wherein the step of 【9】 comprises: the iterative optimization is performed by taking the loss function value as an iterative index of model training, and the hyperparameters are optimized according to the iterative results.

10. The method according to claim 1, wherein the step of 【11】 further comprises: after the evaluation is completed, the evaluation model and the evaluation results are evaluated by using statistical indexes. ​ ​ ​ ​ ​ ​ ​

Images