A method and system for predicting regional atmospheric pollutant distribution

Abstract

The application discloses a kind of regional atmospheric pollutant distribution prediction method and system, comprising: establishing calculation domain to be predicted region, and carrying out grid division;With the data of pollutant measuring point and meteorological measuring point and pollutant emission source intensity as known condition, non-steady-state numerical simulation is carried out, and the data set of global variable on time sequence is calculated;According to the data set of global variable on time sequence, neural network group for pollutant distribution prediction is established and trained;According to the data collected by pollutant measuring point and meteorological measuring point, the pollutant distribution prediction in the region to be predicted is carried out by neural network group.The application carries out a large number of calculations through the numerical simulation of first step, and the grid scale can be much smaller than the grid scale of other commonly used numerical calculations, to provide sufficient and accurate data set for the machine learning prediction of second step, and the data amount and spatial resolution are significantly higher than the existing discrete measurement point method.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of environmental protection, in particular to a method and system for predicting the distribution of regional air pollutants. BACKGROUND

[0002] The state has strict requirements for pollutant emissions of different factories and systems. The local regulatory departments mainly supervise the real-time network of the pollutant measurement values of production enterprises. In addition, for regulatory purposes, regulatory agencies or enterprise environmental protection departments can also install measurement points at other locations to further regulate factory production behavior and supervise the pollutant emissions of production enterprises. However, the distribution of these measurement points is discrete and sparse, which is not sufficient to represent the overall regional pollutant distribution. On the other hand, especially for industrial areas where production enterprises are relatively concentrated, it is also necessary to accurately define the factory responsibility range. At the same time, the influence of weather, climate, etc. on the diffusion and evolution of pollutants should be considered to more accurately lock the direct or indirect sources of pollutants and issue control or warning instructions. The basis of pollutant concentration prediction is wind environment analysis or meteorological analysis. Existing technologies mainly include the following two types: numerical simulation methods of different scales, but these methods often have problems such as difficulty in model establishment, poor calculation accuracy, and low simulation efficiency. Machine learning methods based on sufficient databases, but they need a large amount of accurate data to support learning and training, and the existing number of measurement points is not sufficient to provide sufficient accurate data.

[0003] A method for predicting atmospheric pollutant concentration based on WRF-Chem is disclosed in Chinese patent literature, with publication number CN118013769B and publication date 2024-06-14, which includes the following steps: S1, WRF-Chem parameterization scheme selection: obtain meteorological data, topographic data and pollution source list of the target area, and simulate the air pollutant concentration simulation value of each grid point in the region through WRF-Chem mode; S2, WRF-Chem simulation result correction: use the observation data of the observation station to correct the air pollutant concentration simulation value of each grid point to obtain the correction coefficient; S3, obtain the future prediction data of air pollutants in the target area through the NCEP global numerical weather prediction mode GFS data, and based on the correction coefficient obtained in step S2, realize the prediction of the atmospheric pollution concentration at any position in the target area. However, this technology is simply a numerical simulation simulation method, which is highly professional, the model is complex and difficult to establish, and the simulation grid scale is large, which will cause large errors between the calculation results and the actual results, but pursuing high precision will make the solving process extremely long and low in simulation efficiency. SUMMARY

[0004] The present application is to overcome the problems of difficulty in model establishment, poor calculation accuracy and low simulation efficiency in the prior art by using a numerical simulation method, and lack of sufficient number of measurement points to provide a large amount of accurate data for learning and training by using a machine learning method, and provides a regional atmospheric pollutant distribution prediction method and system.

[0005] In order to achieve the above-mentioned purpose, the present application adopts the following technical solutions:

[0006] A regional atmospheric pollutant distribution prediction method comprises:

[0007] A calculation domain is established for a region to be predicted, and grid division is performed;

[0008] Data of pollutant measurement points and meteorological measurement points and pollutant emission source intensity are taken as known conditions, and non-steady-state numerical simulation is performed to obtain a data set of global variables in time sequence;

[0009] According to the data set of global variables in time sequence, a neural network group for pollutant distribution prediction is established and trained;

[0010] According to the data collected by the pollutant measurement points and the meteorological measurement points, the neural network group is used to predict the pollutant distribution in the region to be predicted.

[0011] In the present application, the data correlation relationship based on the mechanism is analyzed first, and the analysis result is used to guide the establishment of the neural network group for pollutant prediction. In the specific field of pollutant diffusion and evolution in the region, the prior theoretical knowledge in fluid mechanics and chemical reaction is used for the structure design of the neural network group. In the same quality database, the reliability is better than that of the general neural network, and the interpretability problem of machine learning is improved. A large amount of calculation is performed through the first step of numerical simulation, and the grid scale can be much smaller than that of other commonly used numerical calculations, so as to provide sufficient and accurate data set for the second step of machine learning prediction. The data amount and spatial resolution are obviously higher than those of the existing discrete measurement point method, and the data acquisition cost is also much lower than that of using fixed measurement and unmanned aerial vehicle patrol to collect data.

[0012] As a preferred, the calculation domain is established for the region to be predicted, and the grid division comprises:

[0013] The range of the region to be predicted is determined according to the object of interest, the three-dimensional space range of the object of interest and its content are obtained, the space region after simplification is taken as the calculation domain;

[0014] The calculation domain is divided into a plurality of grids, each grid is empty, or includes at least one of the pollutant measurement points, the meteorological measurement points and the pollutant emission source.

[0015] As preferred, the neural network group comprises: a distribution sub-network converting the numerical simulation result into a variable distribution field; a pollution source sub-network taking the output of the distribution sub-network and the numerical simulation result as input and outputting the influence result of pollution source to the prediction sub-network; a convection sub-network outputting the influence result of pollutant convection to the prediction sub-network; a diffusion sub-network outputting the influence result of pollutant diffusion to the prediction sub-network.

[0016] a prediction sub-network outputting the final pollutant distribution prediction result.

[0017] As preferred, the distribution sub-network takes the environmental data of the meteorological measuring point at a certain time in the numerical simulation result, the concentration data of the pollutant measuring point and the intensity data of the pollutant emission source as input.

[0018] The environmental parameter distribution and the pollutant concentration distribution in each grid at the same time are taken as output to establish a neural network sub-model and train it.

[0019] As preferred, in the pollution source sub-network, the output result of the distribution sub-network at the current time is taken as input together with the input of the distribution sub-network at the next time.

[0020] The pollutant concentration increment contributed by the pollution source in each grid is taken as output to establish a neural network sub-model and train it.

[0021] As preferred, in the convection sub-network, the output result of the distribution sub-network at the current time is taken as input together with the environmental data of the meteorological measuring point and the concentration data of the pollutant measuring point at the next time.

[0022] The pollutant concentration increment contributed by the convection with the adjacent grid in each grid is taken as output to establish a neural network sub-model and train it.

[0023] As preferred, in the diffusion sub-network, the output result of the distribution sub-network at the current time is taken as input together with the environmental data of the meteorological measuring point and the concentration data of the pollutant measuring point at the next time.

[0024] The pollutant concentration increment contributed by the mutual diffusion with the adjacent grid in each grid is taken as output to establish a neural network sub-model and train it.

[0025] As preferred, in the prediction sub-network, the output result of the pollution source sub-network is taken as input together with the output results of the diffusion sub-network and the convection sub-network.

[0026] The total amount of the pollutant concentration change in each grid is taken as output to establish a neural network sub-model and train it.

[0027] As preferred, a pollutant concentration threshold value is set, the predicted pollutant concentration in each grid is compared with the pollutant concentration threshold value, and if the predicted pollutant concentration in the grid is greater than the pollutant concentration threshold value, a pollutant emission warning is given for the grid exceeding the pollutant concentration threshold value.

[0028] A prediction system for regional atmospheric pollutant distribution comprises:

[0029] A data acquisition module acquires meteorological observation point data and pollutant observation point data and obtains the intensity of a pollutant emission source; a regional preprocessing module establishes a calculation domain and performs grid division for a region to be predicted;

[0030] A numerical simulation module performs non-steady-state numerical simulation using the data of pollutant observation points and meteorological observation points and the intensity of the pollutant emission source as known conditions to calculate a data set of global variables over time;

[0031] A neural network module establishes and trains a neural network group for pollutant distribution prediction according to the data set of global variables over time.

[0032] The present application has the following beneficial effects: based on the analysis of the mechanism of data correlation, the establishment of the neural network group for pollutant prediction is guided, in the specific field of pollutant diffusion and evolution in a region, the prior theoretical knowledge in fluid mechanics and chemical reactions is used for the structural design of the neural network group, the reliability of the general neural network is achieved in the same quality database, and the interpretability problem of machine learning is improved; a large amount of calculation is performed through the first-step numerical simulation simulation, the grid scale can be much smaller than the grid scale of other commonly used numerical calculations, sufficient and accurate data sets are provided for the second-step machine learning prediction, the data amount and spatial resolution are obviously higher than those of the existing discrete measurement point method, and the data acquisition cost is much lower than that of the fixed measurement and unmanned aerial vehicle flying. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flowchart of a regional atmospheric pollutant distribution prediction method in the present application.

[0034] Figure 2 is a schematic diagram of grid division for a calculation domain in an embodiment of the present application.

[0035] In the figure: 1, grid; 2, pollutant emission source; 3, meteorological observation point; 4, pollutant observation point. DETAILED DESCRIPTION

[0036] The present application will be further described below in combination with the drawings and specific embodiments.

[0037] The basis of pollutant concentration prediction is wind environment analysis or meteorological analysis. The existing technology mainly includes two types: numerical simulation simulation means of different scales, mainly including macro-scale meteorological simulation, mesoscale meteorological simulation, micro-scale wind environment simulation, computational fluid dynamics (CFD) simulation, multi-scale hybrid simulation, etc. However, these methods often have problems such as difficulty in model establishment, poor calculation accuracy, and low simulation efficiency. Based on sufficient database, using machine learning method can analyze the correlation of historical meteorological and pollutant distribution data, and predict the output value under given input conditions, but there are problems of insufficient training data and low calculation efficiency.

[0038] In order to overcome the problems and shortcomings produced by using one of the above two methods alone, the present application combines numerical simulation and machine learning method, uses the results of numerical simulation to provide data for machine learning, and provides a regional atmospheric pollutant distribution prediction method as shown in Figure 1 The method comprises the following steps:

[0039] A calculation domain is established for the region to be predicted, and grid division is performed;

[0040] The data of pollutant measuring points and meteorological measuring points and the pollutant emission source intensity are taken as known conditions, and non-steady-state numerical simulation is performed to calculate the data set of the global variable in time sequence;

[0041] According to the data set of the global variable in time sequence, a neural network group for predicting the distribution of pollutants is established and trained;

[0042] According to the data collected by the pollutant measuring points and the meteorological measuring points, the distribution of pollutants in the region to be predicted is predicted by the neural network group.

[0043] It should be noted that in the present application, the data correlation relationship based on mechanism is first analyzed, and the analysis results are used to guide the establishment of the neural network group for predicting pollutants. In the specific field of pollutant diffusion and evolution in the region, the prior theoretical knowledge in fluid mechanics and chemical reaction is used for the structure design of the neural network group, which achieves better reliability than the general neural network in the same quality database, and improves the interpretability problem of machine learning. A large amount of calculation is performed through the numerical simulation simulation in the first step, and the grid scale can be much smaller than the grid scale of other commonly used numerical calculations, so as to provide sufficient and accurate data set for the machine learning prediction in the second step. The data amount and spatial resolution are significantly higher than the existing discrete measurement point method, and the data acquisition cost is much lower than the fixed measurement and unmanned aerial vehicle flying method.

[0044] As a specific embodiment, the neural network group comprises a distribution sub-network for converting the numerical simulation results into a variable distribution field;

[0045] The output of the distribution sub-network and the numerical simulation result are taken as inputs: the pollution source sub-network outputs the influence result of the pollution source to the prediction sub-network; the convection sub-network outputs the influence result of the convection of the pollutant to the prediction sub-network; and the diffusion sub-network outputs the influence result of the diffusion of the pollutant to the prediction sub-network;

[0046] The prediction sub-network outputs the final prediction result of the distribution of the pollutant.

[0047] It should be noted that the neural network group in the application is composed of five sub-neural networks, and the most basic fully connected neural network can be used for all sub-neural networks to achieve the purpose of the application, and other existing neural network structures can also be selected according to actual needs. In the neural network group, the distribution sub-network is the uppermost neural network, which simulates the acquisition process of the distribution state of the global variable in each grid in the numerical simulation process; the pollution source sub-network, the convection sub-network and the diffusion sub-network are the intermediate layer neural networks, which obtain the influence result of different influence factors on the distribution of the pollutant according to the distribution state of the global variable in each grid; and the prediction sub-network is the lowermost neural network, which comprehensively predicts the final distribution of the pollutant according to the output result of the intermediate layer neural network. All sub-neural networks are trained by the data set of the numerical simulation result, and finally form a neural network group capable of predicting the concentration of the pollutant.

[0048] It should be noted that the application object of the application is a nonlinear extremely strong flow reaction problem, and according to the control equation describing the detailed mechanism, the problem is divided into three different influence factors of convection, diffusion and source, and the sub-neural networks are trained respectively to form a neural network group to calculate and solve the distribution and evolution of the flow field.

[0049] Specifically, for the problem of strong mechanism description y=y1+y2+y3=f(x), the traditional CFD-machine learning coupling method is to solve y directly from x; and the method of the application is to use multiple sub-neural networks to realize the approximate solution of x->y'1, x->y'2 and x->y'3, and then further correct to obtain the final output y. Therefore, compared with the existing technology, the application has the following two advantages:

[0050] Firstly, the sub-neural networks obtained by this splitting method have clear physical meaning, which improves the defect of the traditional CFD-machine learning coupling black box method in interpretability; secondly, because the solution is pushed forward on the grid of the calculation domain, y i =y 1,i +y 2,i +y 3,i (i is the grid number), it is easy to know that the value domain of y that can be predicted by the traditional network on the whole domain is:

[0051] y min ≥y1,min +y 2,min +y 3,min , y max ≤ y 1,max +y 2,max +y 3,max

[0052] And the value range of y that the split network can predict is:

[0053] y min =y 1,min +y 2,min +y 3,min , y max =y 1,max +y 2,max +y 3,max

[0054] Only a simple modification (such as adding a physical constraint: y' i =y' 1,i +y' 2,i +y' 3,i ) to the prediction sub-network is needed to achieve it. Note that this modification only increases the error caused by the strong physical constraint in the prediction sub-network; obviously, such a simple modification is extremely difficult to achieve in a traditional network. In summary, from the perspective of variable value range, this split method has the ability to cover a wider range of parameters, and has a wider range of applications compared to existing technologies.

[0055] The prior knowledge used in the application mainly comes from the basic equations and mechanisms in disciplines such as fluid mechanics, heat and mass transfer, and chemical reactions. The pollutant concentration increment in a grid is decomposed into the contribution of the source, the contribution of convection and diffusion, and a neural network group composed of multiple sub-neural networks is established accordingly. Therefore, using other methods to construct neural networks of different structures, or constructing other machine learning algorithms, can also achieve the purpose of the application; but the algorithms constructed in this way may not have clear physical meaning, i.e. poor interpretability, and it is also difficult to iterate or maintain them.

[0056] As a specific embodiment, as shown in Figure 2 , a calculation domain is established for the region to be predicted, and the grid division includes:

[0057] According to the object of interest, the range of the region to be predicted is determined, the three-dimensional space range and the content of the object of interest are obtained, and the simplified space region is taken as the calculation domain;

[0058] The calculation domain is divided into a plurality of grids 1, each grid is empty, or includes at least one of a pollutant measuring point 4, a meteorological measuring point 3, and a pollutant emission source 2.

[0059] It should be noted that according to the geographic information system (GIS system) topography, building surveying information, the determination of the calculation domain of the to-be-predicted region and the calculation grid division are performed, and the grid information including the node position is stored. The calculation domain and the grid division include the following steps:

[0060] The main body (government department, regulatory unit, etc.) implementing the method can determine the to-be-predicted region range according to the specific object of interest. For example, the environmental protection bureau of the A city B district can take the entire jurisdiction as the object of interest; the head of the environmental protection department of the C factory takes the entire factory as the object of interest.

[0061] The three-dimensional spatial range and its contents of the object of interest are obtained from the GIS and simplified. For example, the outer boundary of the calculation domain is the spatial boundary of the object of interest, the inside of the calculation domain retains large buildings (the outer wall surface of the large building becomes a solid wall boundary in the calculation domain), ignores small buildings (considers that the place is ventilated on four sides and there is no obstacle to block the air flow), etc., and the obtained simplified spatial region is the calculation domain.

[0062] The division method of the calculation domain is consistent with the grid division method in CFD calculation, and the calculation domain is divided into grids with similar geometric sizes, that is, the size of the divided grid is within the allowable size fluctuation range. Generally speaking, the smaller the grid size, the more accurate the result is, under the condition that the computing power is allowed. In actual operation, the grid size is preliminarily determined by using the CFL criterion, and the grid is improved by increasing or decreasing the grid size, locally encrypting the boundary layer grid, etc., to determine that the "grid independence" requirement is met, that is: the accuracy will not be obviously improved by continuing to encrypt the grid. In three-dimensional space, only the space needs to be divided into a plurality of small blocks (cubes can be used, and deformation is allowed).

[0063] The role of the node is to change the continuous space into discrete points, and to approximate the continuous distribution in the space by discrete points. By drawing the grid, the node position can be intuitively seen. Different CFD algorithms have different determination methods for the grid and the node. Commonly used algorithms include: finite volume method (FVM), finite difference method (FDM), finite element method (FEM), meshless method (Meshless Methods), etc. Taking the most classic SIMPLE algorithm of FVM as an example, two sets of nodes are generally used: one set of nodes is placed at the center of the grid, which is used to store scalar quantities such as density; the other set of nodes is placed at the center of the grid surface, which is used to store vectors such as velocity. There are also methods of directly using the grid vertex as the node itself. Each method has different requirements for the node, and for the same calculation object, different grid and node determination methods can also be used, as long as it meets the "grid independence" requirement.

[0064] As a specific embodiment, with the data of the pollution measuring points and the meteorological measuring points and the pollution emission source intensity as known conditions, the non-steady-state numerical simulation is performed to obtain the data set of the global variables in time sequence.

[0065] Specifically, according to the wind speed {V m}, the pressure {P m}, the temperature {T m}, the concentration of the pollution measuring point {C p}, and the pollution emission source intensity {S q}, the non-steady-state computational fluid dynamics (CFD) simulation is performed to obtain the data set of the global variables in time sequence in the calculation domain, including the wind speed V i , the concentration of any component C i,j , the convection term Conv i , the diffusion term Diff i , and the pollution source term S i evolution over time. The source term includes the pollution emission source term and the chemical reaction source term. Wherein i is the grid number, and j is the component number. The subscript m corresponds to the meteorological measuring point sequence number; the subscript p corresponds to the pollution measuring point sequence number; and the subscript q corresponds to the pollution emission source sequence number.

[0066] In the numerical simulation part, the to-be-monitored region is cut and divided according to the discrete and sparse distribution of the pollution measuring points, the meteorological measuring points, and the pollution emission sources in the to-be-predicted region and outside the region. The measuring points can be composed of grid-connected pollution emission sources of enterprise groups such as production enterprises, or other meteorological or pollution measuring points arranged by environmental supervision departments or production enterprises.

[0067] For the measuring points at the pollution emission sources, at least the types, flow, temperature, and pressure of the emitted pollutants should be included (that is, sufficient to make the pollution source emission intensity clear). For the meteorological measuring points, at least the temperature, pressure, and wind speed should be included; and for the pollution measuring points, mainly the pollution concentration data should be collected.

[0068] Regarding the selection of the grid size / resolution: for the grid division in general problems, the grid size needs to be determined by balancing the accuracy requirement and the computing capacity. Because the numerical simulation calculation results of the present application will be included in the pollution distribution database, there is a long-term use purpose, and more computing resources are invested in the numerical simulation part to perform high-resolution grid division to construct a reliable database for the subsequent neural network group to train.

[0069] By incorporating pollutant emission source information from production enterprises into the source terms of the control equations, unsteady-state numerical simulations are performed on the control equations and models for computational fluid dynamics, mass transfer (both necessary), heat transfer, and chemical reaction (the latter two optional) to obtain the distribution patterns and variation laws of variables such as air temperature, wind speed, and pollutant concentration within the monitoring area. For example, in predicting haze (particulate pollutants), Ansys Fluent commercial software is used. Based on the aforementioned determined computational domain, mesh generation, and boundary conditions, the Eulerian model is selected for fluid flow solution, and the DPM (Discrete Phase Model) model is selected for particulate flow solution. A particulate event model is added through UDF (User Defined Function) (this step is optional; for example, a kernel-driven collision-aggregation / fragmentation model can be added) to obtain the distribution of particulate pollutants in each computational domain.

[0070] As a specific implementation, in the prediction part of the neural network group, the high-resolution distribution results obtained from the numerical simulation step are used as the data source for machine learning to establish a pollutant distribution database. New pollutant distribution data can be added to the database at any time through numerical simulation calculations using newly acquired data (any change in any point in the meteorological or pollutant monitoring data is considered new). The data in the database is preprocessed to obtain training and testing sets, using three common partitioning methods: hold-out, cross-validation, and bootstrapping. The structure of the neural network group is determined, using the meteorological and pollutant monitoring data from the database as input and the pollutant concentration distribution of each grid within the region as output. Through training and testing, the final neural network group used for pollutant concentration prediction is obtained.

[0071] The distributed subnetwork takes environmental data from meteorological monitoring points at a certain moment in the numerical simulation results, combined with concentration data from pollutant monitoring points and intensity data from pollutant emission sources, as input.

[0072] Using the distribution of environmental parameters and pollutant concentrations in each grid at the same time as the output, a neural network sub-model is established and trained.

[0073] Specifically, the input to the distributed subnetwork is: the wind speed {V} at the meteorological measuring point measured at any time t. m,t Pressure {P} m,t}, temperature {T m,t}、Pollutant monitoring point concentration {C p,t}、Pollutant emission source intensity {S q,t The output of the distributed subnetwork is: the wind speed distribution field {V} at the same time t. i,t}、Pressure distribution field {P i,t Temperature distribution field {T}i,t}, pollutant concentration distribution field {C i,t}. The distribution sub-network is trained using the data set of numerical simulation, so that the trained distribution sub-network can obtain the meteorological data distribution and the pollutant concentration distribution in each grid in the to-be-predicted area according to the input meteorological measurement point data, pollutant measurement point data and pollutant emission source intensity at a certain moment.

[0074] In the pollutant source sub-network, the output result of the current moment distribution sub-network is combined with the input of the next moment distribution sub-network as input;

[0075] The pollutant concentration increment contributed by the pollutant source in each grid is taken as output, and a neural network sub-model is established and trained.

[0076] Specifically, the input of the pollutant source sub-network is: the wind speed distribution field {V i,t}, the pressure distribution field {P i,t}, the temperature distribution field {T i,t}, the pollutant concentration distribution field {C i,t} at any moment t, and the measured meteorological measurement point wind speed {V m,t+dt}, pressure {P m,t+dt}, temperature {T m,t+dt}, pollutant measurement point concentration {C p,t+dt} and pollutant emission source intensity {S q,t+dt} at the next moment t+dt. (When the value at t+dt cannot be obtained, any reasonable value defined by the user can be used, such as an experience estimate, a simulation value, an extrapolated value obtained according to a trend, or the measured value at t is directly used instead).

[0077] The output of the pollutant sub-network is the pollutant concentration increment C i contributed by the pollutant emission source S a,i , which has a physical meaning of the change of the pollutant concentration in grid i within dt due to the pollutant emission source and the chemical reaction source term. The pollutant source sub-network is trained using the output result of the distribution sub-network at the same moment and the data of the meteorological measurement point, the pollutant measurement point and the pollutant emission source at the next moment, to obtain a pollutant source sub-network meeting the requirements of the neural network group.

[0078] In the convection sub-network, the output result of the current moment distribution sub-network is combined with the environmental data of the meteorological measurement point and the concentration data of the pollutant measurement point at the next moment as input;

[0079] The pollutant concentration increment contributed by the convection with the adjacent grid in each grid is taken as output, and a neural network sub-model is established and trained.

[0080] Specifically, the input of the convection sub-network is: the wind speed distribution field {V i,t}, the pressure distribution field {P i,t}, the temperature distribution field {T i,t}, the pollutant concentration distribution field {C i,t} at any moment t, and the measured meteorological measurement point wind speed {V m,t+dt}, pressure {P m,t+dt}, temperature {T m,t+dt}, pollutant measurement point concentration {C p,t+dt} and pollutant emission source intensity {S q,t+dt} at the next moment t+dt. (When the value at t+dt cannot be obtained, any reasonable value defined by the user can be used, such as an experience estimate, a simulation value, an extrapolated value obtained according to a trend, or the measured value at t is directly used instead).i,t}, pressure distribution field {P i,t}, temperature distribution field {T i,t}, pollutant concentration distribution field {C i,t}, and the wind speed of the meteorological measuring point {V m,t+dt}, pressure {P m,t+dt}, temperature {T m,t+dt}, and pollutant concentration of the measuring point {C p,t+dt} obtained at the next moment t+dt. (When the value at the moment t+dt cannot be obtained, any reasonable value defined by the user can be used, such as an empirical estimate, a simulation value, an extrapolated value obtained according to a trend, or directly using the value measured at the moment t instead.)

[0081] The output of the convection sub-network is: Conv i and Conv in The physical meaning of the pollutant concentration increment C b,i contributed by the grid i is the change in pollutant concentration of the grid i caused by convection in dt. The subscript in represents i-neighbor, representing all grids adjacent to the grid i. The convection sub-network is trained using the output results of the distribution sub-network at the same moment and the simulation results at the next moment, to obtain a convection sub-network that meets the requirements of the neural network group.

[0082] In the diffusion sub-network, the output results of the distribution sub-network at the current moment are used as input, combined with the environmental data of the meteorological measuring point and the concentration data of the pollutant measuring point at the next moment.

[0083] The pollutant concentration increment contributed by the mutual diffusion between each grid and its adjacent grids is taken as the output, and a neural network sub-model is established and trained.

[0084] Specifically, the input of the diffusion sub-network is: the wind speed distribution field {V i,t}, pressure distribution field {P i,t}, temperature distribution field {T i,t}, and pollutant concentration distribution field {C i,t} at any moment t, and the wind speed of the meteorological measuring point {V m,t+dt}, pressure {P m,t+dt}, temperature {T m,t+dt}, and pollutant concentration of the measuring point {C p,t+dt} obtained at the next moment t+dt. (When the value at the moment t+dt cannot be obtained, any reasonable value defined by the user can be used, such as an empirical estimate, a simulation value, an extrapolated value obtained according to a trend, or directly using the value measured at the moment t instead.)

[0085] The output of the diffusion sub-network is: Diff i and Diff inThe pollutant concentration increment C contributed by the pollutant source c,i The physical meaning is the pollutant concentration change of the grid i within dt caused by diffusion. The subscript in means i-neighbor, representing all grids adjacent to grid i. The diffusion subnetwork is trained using the output results of the same time distribution subnetwork and the results of the numerical simulation at the next time, to obtain a diffusion subnetwork meeting the requirements of the neural network group.

[0086] In the prediction subnetwork, the output results of the pollutant source subnetwork are combined with the output results of the diffusion subnetwork and the convection subnetwork as inputs;

[0087] The total amount of pollutant concentration change in each grid is taken as the output, and a neural network submodel is established and trained.

[0088] Specifically, the inputs of the prediction subnetwork are: the pollutant emission source S at any time t i The pollutant concentration increment C contributed by the pollutant source a,i Conv i Conv in The pollutant concentration increment C contributed by the pollutant source b,i Diff i Diff in The pollutant concentration increment C contributed by the pollutant source c,i The output of the prediction subnetwork is: C total,i The physical meaning is the total amount of pollutant concentration change of grid i within dt. The pollutant concentration prediction result at the next time t+dt can be obtained by combining the pollutant concentration distribution at any time t with the total amount of pollutant concentration change of each grid within dt. Other outputs of the prediction subnetwork include the environmental parameter distribution in each grid; the neural network group obtained by combining the five subnetworks can predict the pollutant concentration of each grid.

[0089] After training the neural network group formed by the combination of the five subnetworks, when predicting the pollutant concentration distribution, the meteorological measurement point wind speed {V m}, pressure {P m}, temperature {T m}, and pollutant measurement point concentration {C m} can be directly input into the neural network group for result prediction.

[0090] Optionally, when predicting the pollutant concentration distribution, only the distribution subnetwork needs to be used once at t=0 to obtain the distribution at t=0 (i.e., to obtain a relatively accurate initial distribution field) as the starting input of other subnetworks. For example, the meteorological measurement point wind speed {V m,t0}, pressure {P m,t0}, temperature {T m,t0} pollutant concentration of measuring point {C p,t0} pollutant emission source intensity {S q,t0} as the input of the distribution sub-network. The output of the distribution sub-network is: the wind speed distribution field {V i,t0} at the initial time t0, the pressure distribution field {P i,t0} at the initial time t0, the temperature distribution field {T i,t0} at the initial time t0, and the pollutant concentration distribution field {C i,t0} at the initial time t0. Then, the result prediction at the time t1=t0+dt is carried out by the subsequent sub-network. The result prediction at the time t1=t0+dt can be input into the intermediate layer sub-network and the lowermost layer sub-network as new input data to carry out the result prediction at the time t2=t1+dt, and the prediction at the time t n =t n-1 +dt can be continuously carried out only by using the intermediate layer sub-network and the lowermost layer sub-network.

[0091] As an optional embodiment, a pollutant concentration threshold is set, and the predicted pollutant concentration in each grid is compared with the pollutant concentration threshold. If the predicted pollutant concentration in the grid is greater than the pollutant concentration threshold, a pollutant emission warning is given to the grid exceeding the pollutant concentration threshold.

[0092] Specifically, the current pollutant limit requirement can be taken as a standard (which can be a national standard, a local standard, or an arbitrary reasonable value artificially defined by a competent department). In the simulation and prediction of pollutant distribution, it is judged according to the above standard whether there is a grid with pollutant concentration exceeding the limit in the region at present or in the future period, and a pollutant emission warning is given to the grid exceeding the limit.

[0093] In addition to the prediction method of regional atmospheric pollutant distribution, the present application also provides a prediction system of regional atmospheric pollutant distribution, comprising:

[0094] A data acquisition module is configured to acquire meteorological measuring point data and pollutant measuring point data, and obtain the intensity of a pollutant emission source. A regional preprocessing module is configured to establish a calculation domain and divide a grid for a region to be predicted.

[0095] A numerical simulation module is configured to perform non-steady-state numerical simulation by taking the data of the pollutant measuring point and the meteorological measuring point and the intensity of the pollutant emission source as known conditions, and calculate a data set of global variables over time.

[0096] A neural network module is configured to establish and train a neural network group for pollutant distribution prediction according to the data set of global variables over time.

[0097] A database is configured to save the results calculated by the numerical simulation module.

[0098] It should be noted that the existing machine learning prediction method mainly analyzes and studies the historical monitoring data, mines the relationship between the data, finds the pollution rule, and establishes a prediction model. The meteorological data used for analysis by this method comes from the public or paid databases of domestic and foreign meteorological institutions (low grid resolution, scale generally kilometer level or above), and the pollutant distribution data comes from limited measuring points. That is, it is difficult to provide sufficient and detailed data for accurate statistical prediction under existing conditions. This purely black box method lacks interpretability and cannot reveal the natural scientific laws in the process. It is difficult to extrapolate. That is, it is difficult to predict new situations beyond the historical data range. In summary, the purely statistical prediction method cannot meet the requirements of theoretical analysis in frontier research and change response in practical application.

[0099] The statistical prediction method needs sufficient and accurate data as input, while the objective actual extremely sparse pollutant concentration measuring point distribution is difficult to meet the requirement of accurate prediction for any point in the whole region. The present application can provide sufficient and accurate data set for the second step of machine learning prediction through the first step of numerical simulation CFD algorithm. The data amount and spatial resolution are obviously higher than the highly discrete measurement method, and the data acquisition cost is much lower than using fixed measurement, unmanned aerial vehicle patrol and other methods to collect data.

[0100] The existing numerical simulation method is based on emission source data, and calculates the evolution process of pollutant distribution from the theoretical point of view of chemistry and atmospheric physics, fluid mechanics, etc., to analyze the change rule of air quality. The problems of this method are: large scale, coarse grid, difficult modeling. Specifically: the existing model solving grid generally has low resolution (mostly kilometer scale), and often ignores geographical factors such as terrain, buildings, etc.; combined with the reasons that the existing pollutant concentration measuring point distribution is sparse, and the atmospheric movement itself is very complex, etc., the simulation result has large error compared with the actual result (the above factors are not purely listed, but are interrelated). If you want to implement fine modeling and solving by reducing the grid scale and increasing the measuring points, the calculation efficiency will be significantly reduced. In summary, the purely numerical simulation method cannot meet the time efficiency and cost requirements in practical application.

[0101] In the present application, a large amount of computing power is invested in the early stage, and a database is constructed by solving a fine mathematical model, which is used for machine learning training and prediction in the later stage. In other words, because the results obtained by numerical simulation are put into the database for long-term use, the reasonable (even high) calculation cost is acceptable for the present application after the initial large amount of computing power is evenly distributed in the long term. Therefore, the grid scale of the numerical simulation method in the present application can be much smaller than the grid scale of other commonly used numerical simulation methods (kilometer level), and various necessary mechanisms (such as particle agglomeration and secondary pollutant reaction) can be added to make the model complete and improve the solving accuracy comprehensively.

[0102] The above examples are further elaboration and illustration of the present application, so as to be understood, and are not any limitation of the present application, any modification, equivalent replacement and improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method of predicting the distribution of an atmospheric pollutant in a region, characterized by, The method comprises the following steps: a calculation domain is established for the region to be predicted, and the calculation domain is divided into grids; With the data of the pollution measuring points and the weather measuring points and the pollution emission source intensity as known conditions, non-steady-state numerical simulation is performed to obtain a data set of the global variables over time, including the wind speed, pressure, temperature, concentration of any component, and the evolution of the convection term Conv i , the diffusion term Diff i , and the pollution source term S i of each control equation over time for any grid i; a neural network group for predicting pollutant distribution is established and trained according to the data set; in the neural network group, a distribution sub-network simulates the process of obtaining the distribution state of the global variable in each grid in the numerical simulation process; the output O of the current time distribution subnetwork t the input I of the next time distribution subnetwork t+1 as input; the pollutant concentration increment contributed by the pollution source in each grid as output to the prediction subnetwork; The convection subnetwork takes as input the environmental data H t at the weather station at the next time instant t+1 and the concentration data W t+1 at the pollution station; and outputs as output to the prediction subnetwork the increments of the concentration of the pollutant in each grid due to the contribution of the adjacent grids by convection. The diffusion subnetwork takes as input the concentration of pollutants in each grid cell at time t and outputs the concentration of pollutants in each grid cell at time t+1. t , H t+1 , W t+1 The concentration of pollutants in each grid cell at time t is the input to the prediction subnetwork. The output of the prediction subnetwork is the concentration of pollutants in each grid cell at time t+1. a prediction sub-network outputs the prediction result of the pollutant distribution; According to the data collected by the pollution measuring point and the weather measuring point, the distribution of the pollutants in the region to be predicted is predicted through a neural network group. When predicting, I t+1 , H t+1 , W t+1 Use custom reasonable values.

2. The method of claim 1, wherein the method further comprises: the step of establishing the calculation domain for the region to be predicted and dividing the calculation domain into grids comprises the following steps: the range of the region to be predicted is determined according to the object of interest, the three-dimensional space range of the object of interest and the content thereof are obtained, and the space region after simplification is taken as the calculation domain; the calculation domain is divided into a plurality of grids, and each grid is empty or includes at least one of a pollutant measuring point, a meteorological measuring point and a pollutant emission source.

3. The method of claim 1, wherein the method further comprises: the distribution sub-network takes the environmental data of the meteorological measuring point at a certain moment, the concentration data of the pollutant measuring point and the intensity data of the pollutant emission source as inputs; in the data set, the environmental parameter distribution and the pollutant concentration distribution in each grid at the same moment are taken as outputs, and a neural network sub-model is established and trained.

4. The method of claim 3, wherein the step of determining the distribution of the atmospheric pollutant in the region is performed by solving a partial differential equation. input I of the next time distribution subnetwork t+1 Custom reasonable values of I include empirical estimates, simulation values, trend extrapolations, or measurements at the current time.

5. The method of claim 3, wherein the step of predicting the distribution of the regional atmospheric pollutant is performed by using a numerical model. environmental data H of the next time instant of the weather measuring point t+1 and concentration data W of the pollutant measuring point t+1 The custom reasonable values of H and W include empirical estimates, simulation values, trend extrapolation values, or measured values at the current time instant.

6. A method of predicting the distribution of an atmospheric pollutant according to any one of claims 1 to 5, wherein, in the prediction sub-network, the output result of the pollutant source sub-network is taken as an input together with the output results of the diffusion sub-network and the convection sub-network; in the data set, the total amount of change of the pollutant concentration in each grid is taken as an output, and a neural network sub-model is established and trained. Combined with the pollutant concentration distribution at any time t dt The total change in pollutant concentration in each grid within the grid can be used to obtain the next time step. t+dt The predicted pollutant concentration results.

7. The method of claim 6, wherein the method further comprises: a pollutant concentration threshold is set, the predicted pollutant concentration in each grid is compared with the pollutant concentration threshold, and if the predicted pollutant concentration in the grid is greater than the pollutant concentration threshold, a pollutant emission warning is given for the grid exceeding the pollutant concentration threshold.

8. A system for predicting the distribution of an atmospheric pollutant in a region, adapted for use in the prediction method of any one of claims 1 to 7, characterised in that, The method comprises the following steps: a data acquisition module acquires meteorological measuring point data and pollutant measuring point data, and obtains the intensity of a pollutant emission source; a region preprocessing module establishes a calculation domain for the region to be predicted and divides the calculation domain into grids; a numerical simulation module performs non-steady-state numerical simulation by taking the data of the pollutant measuring point and the meteorological measuring point and the intensity of the pollutant emission source as known conditions, and calculates a data set of the global variable over time; a neural network module establishes and trains a neural network group for predicting pollutant distribution according to the data set of the global variable over time.

9. The system for predicting the distribution of regional atmospheric pollutants according to claim 8, wherein, a database is further included, which stores the data set of the global variable over time calculated by the numerical simulation module.

Images