CN104881546A - Method for improving prediction efficiency of atmospheric pollution model - Google Patents

Abstract

The invention provides a method for improving the prediction efficiency of an atmospheric pollution model. The method comprises the following steps: pre-processing original meteorological data, acquiring boundary layer parameter data and profile data, acquiring a pollution concentration value of each prediction grid point in the atmospheric pollution model, acquiring a contour line for concerned pollutant concentration according to the pollution concentration values, and acquiring pollution influence area according to the contour line; establishing a polynomial prediction model with a prediction effect by utilizing the boundary layer parameter data, the profile data and the pollution influence area; by using real-time meteorological data in the presence of atmospheric pollution as input, acquiring prediction pollution area according to the polynomial prediction model, importing the prediction pollution area and the position information of a pollution source into an AERMAP module of the atmospheric pollution model to serve as a basis for selecting an evaluation pollution region in the atmospheric pollution model. According to the method, the unnecessary calculation consumption of the grid points in an AERMOD model is reduced, and the prediction efficiency is improved.

Description

A kind of method improving air pollution model forecasting efficiency

Technical field

The present invention relates to safety and the urgent technique field of sudden accident, particularly relate to air pollution diffusion technical field, be specially a kind of method improving air pollution model forecasting efficiency.

Background technology

Along with expanding economy, the impact of industrial pollution source on atmospheric environment is increasing, causes serious environmental disruption, brings serious threat to the living environment of the people.But, industry spot is polluted carry out real-time dynamic monitoring, prediction has again very large difficulty, is not very feasible, therefore, is carried out the very general of the diffusion application of simulation and forecast atmospheric pollution by Air Pollution Diffusion Model.Model of atmospheric diffusion is a kind of environmental model of spatio temporal composite, for describing air, pollutant defeated is moved, spread and diluting effect, it is the powerful measure of carrying out environmental evaluation and environmental forecasting, is assessment spillage risk degree and possibility leakage rate, the best approach analyzing toxic chemical substance atmospheric air leakage accident environmental impact.Near-earth Air Dispersion Modeling is just developed before and after the thirties in 20th century, has developed numerous model of atmospheric diffusion so far.

From 1 day April in 2009, the current law pattern of China comprised estimation mode, AERMOD, ADMS and CALPUFF pattern.The middle and later periods nineties 20th century; associating USEPA of meteorology institute of the U.S. sets up regulation model and improves the council based on up-to-date atmospheric boundary layer and Atmospheric diffusion theory; successfully develop AERMOD model of atmospheric diffusion, instead of ISC model originally with the method scale that a generation is new.AERMOD model is stable state maturity degree, and range of value is within 50km, is starting point with the statistical theory of diffusion, supposes the CONCENTRATION DISTRIBUTION Gaussian distributed to a certain extent of pollutant.This model is made up of three parts: AERMET module (weather data pretreater), AERMAP module (terrain data pretreater) and nucleus module AERMOD (diffusion model).Modular system can be used for the discharge of multiple emission source (comprising point source, source, face and body source), is also applicable to the Simulation and Prediction of rural environments and situation such as multiple discharge diffusion such as urban environment, subdued topography and complicated landform, ground-level source and elevated source etc.Wherein, within this model scope of application is generally 50km, and the landform altitude scope in using must be greater than the scope in selected evaluation and test region.The deficiency of model is: in reality get DEM altitude figures evaluation and test regional extent with pollute compared with the actual area (or evaluation and test region of reality) that causes excessive, result in larger calculated amount, reduce counting yield.Before this, predicted application is all carried out in strict accordance with the design of master pattern to the application of model in all contamination prediction aspects that relates to, and not clear and definite foundation as choosing of evaluation and test region.Domestic and international in the prediction of forest fires, rainfall at present, according to a large amount of weather datas, use the non-linear regression in machine learning and support vector machine technology, obtain forecast model, got good effect.This patent is conceived to: model itself does not have according to historical data in advance by predicting the function chosen landform altitude dem data scope and choose, therefore the innovation of this patent is also just on the basis of master pattern, be subject to the inspiration of correlation predictive method, use the multinomial homing method in the middle of machine learning, trained by a large amount of weather data, set up the multinomial model choosing evaluation and test area size, improving in the past to evaluating and testing regional choice according to indefinite drawback, improving the prediction and calculation efficiency of model.

China commonly uses model of atmospheric diffusion AERMOD Introduction on Principle: the boundary layer parameters data of AERMET and profile data can be determined by the site observation date inputted, or is generated by the National Meteorological Bureau's routine meteorological data inputted (ground data, sounding data).AERMET profile data and the control documents of boundary layer profile data in AERMOD are quoted and are entered AERMOD system, calculate similar parameter, and carry out interpolation to boundary layer profile data.AERMOD is by mean wind speed, level to and vertical amount of turbulence pulsation, thermograde, megadyne temperature, the input such as horizontal Lagrange time scale dispersal pattern, and calculate AERMAP be simplify and standardized A ERMOD landform input data terrain pre processing device, it is by the location parameter (x of each net point of input, y, and Terrain Elevation parameter (xt z), yt, zt) through calculating the terrain data changing into AERMOD data processing, include each grid point locations parameter (x, y, and significant height value z), these data be used for barrier ambient atmosphere diffusion calculating and in conjunction with the isoparametric distribution of wind speed u, thus the distribution can carrying out pollutant levels calculates.The concentration prediction of AERMOD calculates.

1) subdued topography condition (namely not considering the influence of topography)

$C(X,Y,Z)=\frac{Q}{\stackrel{\‾}{u}}{p}_y(y,x)p_z(z,x)$

In formula: Q is discharge of pollutant sources speed, u is effective wind speed, p _y, p _yfor the probability density function of horizontal direction and vertical direction CONCENTRATION DISTRIBUTION.

2) topographic condition is considered:

C _r(x,y,z)＝f.C(x _r,y _r,z _r)+(1-f).C(x _r,y _r,z _p)

In formula: C (x _r, y _r, z _r) be the contribution of horizontal plume, C (x _r, y _r, z _p) contribute for the plume of the influence of topography.

3) concentration prediction in convective boundary layer calculates

The ground concentration predicted in convective boundary layer is made up of the concentration of following three kinds of pollution source:

Part I: because down draft is directly diffused into the direct source on ground:

$C_d\{x_r,y_r,z_r\}=\frac{Q\·f_p}{\sqrt{2\mathrm{\πu}}}\·F_y\·\underset{j=1}{\overset{2}{\mathrm{\Σ}}}\underset{m=0}{\overset{\∞}{\mathrm{\Σ}}}\frac{{\mathrm{\λ}}_j}{{\mathrm{\σ}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-{\mathrm{\ψ}}_{\mathrm{dj}}-{2\mathrm{mz}}_i)}^2}{{{\mathrm{z\σ}}_{\mathrm{zj}}}^2}]+[\frac{{(z-{\mathrm{\ψ}}_{\mathrm{dj}}-2{\mathrm{mz}}_i)}^2}{{{2\mathrm{\σ}}_{\mathrm{zj}}}^2}\left]\right]$

$F_y=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_y}\exp \left[\frac{{-y}^2}{{{2\mathrm{\σ}}_y}^2}\right]$

Part II: because updraft is diffused into the indirect source on mixolimnion top:

$C_d\{x_r,y_r,z_r\}=\frac{Q\·f_p}{\sqrt{2\mathrm{\πu}}}\·F_y\·\underset{j=1}{\overset{2}{\mathrm{\Σ}}}\underset{m=1}{\overset{\∞}{\mathrm{\Σ}}}\frac{{\mathrm{\λ}}_j}{{\mathrm{\σ}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-{\mathrm{\ψ}}_{\mathrm{rj}}-{2\mathrm{mz}}_i)}^2}{{{\mathrm{z\σ}}_{\mathrm{zj}}}^2}]+[\frac{{(z-{\mathrm{\ψ}}_{\mathrm{rj}}-2{\mathrm{mz}}_i)}^2}{{{2\mathrm{\σ}}_{\mathrm{zj}}}^2}\left]\right]$

Part III: penetrate into the plume in mixolimnion upper stabilizer layer, through also reentering mixolimnion after a period of time, and is diffused into ground, namely penetrates source:

$C_d\{x_r,y_r,z_r\}=\frac{Q\·\left({1-f}_p\right)}{\sqrt{2\mathrm{\πu}}}\·F_y\·\mathrm{\Σ}\underset{m=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\frac{{\mathrm{\λ}}_j}{{\mathrm{\σ}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-h_{\mathrm{ep}}-{2\mathrm{mz}}_{\mathrm{ieff}})}^2}{{{\mathrm{z\σ}}_{\mathrm{zp}}}^2}]+[\frac{{(z-h_{\mathrm{ep}}-2{\mathrm{mz}}_{\mathrm{ieff}})}^2}{{{2\mathrm{\σ}}_{\mathrm{zp}}}^2}\left]\right]$

With in above formula, F _yfor horizontal distribution function, σ _yfor Horizontal Diffusion Parameter, z _ifor mixing height, σ _zjfor direct source vertical diffusion parameter, σ _zpfor penetrating source vertical diffusion parameter, ψ _djfor direct source plume overall height, ψ _rjfor indirect source plume overall height, λ _jbe the weight coefficient of rising and sinking two parts plume, m is order of reflection, h _epfor penetrating source height, z _iefffor reflecting surface height in resistant strata.

4) concentration prediction in stable concave surface calculates

On subdued topography, the form being applicable to stable condition AERMOD concentration expression formula is Gauss model:

$C(x,y,z)=\frac{Q}{u}\×F_z\left(x_z{h}_p\right)\×F_y\left(y\right)$

$F_y=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_y}\exp \left[\frac{{-y}^2}{{{2\mathrm{\σ}}_y}^2}\right]$

$F_z=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_z}\underset{n=-\∞}{\overset{\∞}{\mathrm{\Σ}}}\exp \left[\frac{{(z-h_p+2{\mathrm{nh}}_a)}^2}{{{2\mathrm{\σ}}_z}^2}\right]+\exp [\frac{{(z+h_p+{2\mathrm{nh}}_a)}^2}{{{2\mathrm{\σ}}_z}^2}]$

In formula: h _pfor plume height, H _afor vertically mixing limiting altitude, H _pfor emission source height and plume rise height sum, F _yfor the horizontal distribution function of meander.

Model of atmospheric diffusion AERMOD is that stress and strain model is carried out in the evaluation and test region that large DEM landform altitude scope of choosing comprises as far as possible within the effective range of model in actual applications, to reach the object of the coverage comprising actual contamination accident, then on this evaluation and test region, stress and strain model is carried out, because the evaluation and test region chosen is often large than actual pollution range, the net point outside actual pollution range is so just made also to take part in the middle of the computation process of AERMOD model, do the problem of the calculated amount adding the actual region be not affected in selected evaluation and test regional extent undoubtedly like this, add unnecessary calculating, reduce the forecasting efficiency of model.

Summary of the invention

The shortcoming of prior art in view of the above, the object of the present invention is to provide a kind of method improving air pollution model forecasting efficiency, for solving the problem that in prior art, air pollution model calculated amount is large and forecasting efficiency is low.

For achieving the above object and other relevant objects, the invention provides a kind of method improving air pollution model forecasting efficiency, be applied in the process by air pollution model evaluation and test Polluted area, described method comprises: carry out pre-service to original weather data and obtain boundary layer parameters data and profile data, according to described boundary layer parameters data, described profile data, pollution source parameter and terrain data obtain the pollution concentration value of each predicted grid point in air pollution model, obtain the isoline of concern pollutant levels according to described pollution concentration value and obtain pollution effect area according to described isoline, described boundary layer parameters data, described profile data and described pollution effect area is utilized to set up the polynomial forecast model with predicting function, with real time meteorological data when there is atmospheric pollution for input, obtain prediction contaminated area according to described polynomial forecast model, the positional information of described prediction contaminated area and pollution source is imported in air pollution model as choosing the foundation evaluating and testing Polluted area in described air pollution model.

Preferably, by carrying out to the pollution concentration value of each predicted grid point described the isoline that interpolation arithmetic obtains described concern pollutant levels.

Preferably, described acquisition pollution effect area is specially: obtain the distance farthest between 2 in described isoline, contain described isoline at interior square with described distance for the length of side builds, obtain described foursquare area and using described foursquare area as described pollution effect area.

Preferably, described isoline is at least two.

Preferably, utilize described boundary layer parameters data, described profile data and described pollution effect area to set up the polynomial forecast model with predicting function to be specially: with described boundary layer parameters data and described profile data for training data, with described pollution effect area for dependent variable data, through obtaining multinomial model to the repeatedly training of described training data, more described multinomial model is optimized to the polynomial forecast model described in acquisition with predicting function according to square error and the goodness of fit.

Preferably, the polynomial forecast model described in predicting function is:

f(x ₁,x ₂,x ₃...x _m)＝β ₀x ₁ ⁿ+β ₁x ₂ ⁿ+...+β _n-1x _n ⁿ+β _nx ₁ ^n-1x ₂...x _m；

Wherein, x ₁, x ₂, x ₃... x _mthe training input data of the polynomial forecast model that the principal component parameter being respectively boundary layer parameters data and profile data is formed; β ₀, β ₁... β _n-1, β _nbe respectively the coefficient of training input data in polynomial forecast model; N is the number of model training input parameter; M is the number of group item parameter.

Preferably, according to square error and the goodness of fit described multinomial model is optimized and is specially: obtain the square error of model prediction area and and the goodness of fit after, utilize gradient descent method to obtain cost function value of coefficient in polynomial forecast model when minimum value.

Preferably, described model prediction area square error and be: wherein, S _ifor model is to the predicted value of area, S _rifor comprising the foursquare actual value of isoline, n is sample number, and i is i-th sample in sample; The described goodness of fit is: wherein, r ²for institute's established model is to the goodness of fit of problem matching, S _ifor model is to the predicted value of area, for the mean value of model prediction area, x _ifor the training input data of the polynomial forecast model that the principal component parameter of boundary layer parameters data and profile data is formed; for x _imean value; Described cost function is: wherein, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model, r ²for institute's established model is to the goodness of fit of problem matching, S ⁽ⁱ⁾be the area actual value of i-th sample, f (X ⁽ⁱ⁾) be the predicted value of model to area of i-th sample; The formula that gradient descent method adopts is: wherein, α is learning rate, β _jfor the coefficient coefficient of polynomial forecast model training input data, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model.

As mentioned above, a kind of method improving air pollution model forecasting efficiency of the present invention, has following beneficial effect:

1, the present invention combines with model of atmospheric diffusion AERMO, proposes the method that optimum option carries out the evaluation and test region of stress and strain model, decreases the calculating consumption of unnecessary net point, improve forecasting efficiency.

2, the method that the present invention proposes has predictability, reasonably can predict, reduce the blindness that evaluation and test region chosen by AERMOD model in the past of knowing clearly according to the true weather data before accident generation.

3, the method that the present invention proposes has good applicability, can play reference role in good actual scene, the waste of manpower and materials in reduction accident in the middle of actual contamination accident to emergency management and rescue.

Accompanying drawing explanation

Fig. 1 is shown as a kind of schematic flow sheet improving the method for air pollution model forecasting efficiency of the present invention.

Fig. 2 is shown as a kind of schematic diagram improving original model of atmospheric diffusion AERMOD model in the method for air pollution model forecasting efficiency of the present invention.

Fig. 3 is shown as and of the present inventionly a kind ofly improves in the method for air pollution model forecasting efficiency the schematic diagram choosing new evaluation and test region.

Fig. 4 is shown as the algorithm flow chart that a kind of method improving air pollution model forecasting efficiency of the present invention is optimized original AERMOD model.

Embodiment

Below by way of specific instantiation, embodiments of the present invention are described, those skilled in the art the content disclosed by this instructions can understand other advantages of the present invention and effect easily.The present invention can also be implemented or be applied by embodiments different in addition, and the every details in this instructions also can based on different viewpoints and application, carries out various modification or change not deviating under spirit of the present invention.

The object of the present invention is to provide a kind of method improving air pollution model forecasting efficiency, for the problem solving in prior art, air pollution model calculated amount is large and forecasting efficiency is low.Below by detailed description a kind of square ratio juris and embodiment improving air pollution model forecasting efficiency of the present invention, those skilled in the art are made not need creative work can understand a kind of method improving air pollution model forecasting efficiency of the present invention.

The invention reside in and propose a kind ofly to choose the method for comparatively cutting the evaluation and test region comprising actual accidents coverage close to actual conditions, stress and strain model is carried out in this evaluation and test region, reduce reality not by the calculating of net point in accident impact region, improve the forecasting efficiency of model.

Particularly, in the present embodiment, as shown in Figure 1, the present embodiment provides a kind of method improving air pollution model forecasting efficiency, is applied in the process by air pollution model evaluation and test Polluted area, said method comprising the steps of.

Step S11, pre-service is carried out to original weather data and obtains boundary layer parameters data and profile data, obtain the pollution concentration value of each predicted grid point in air pollution model according to described boundary layer parameters data, described profile data, pollution source parameter and terrain data, obtain the isoline of concern pollutant levels according to described pollution concentration value and obtain pollution effect area according to described isoline.In the present embodiment, model of atmospheric diffusion adopts AERMOD model of atmospheric diffusion, and specifically by carrying out to the pollution concentration value of each predicted grid point described the isoline that interpolation arithmetic obtains described concern pollutant levels, described isoline is at least two.

In the present embodiment, described acquisition pollution effect area is specially: obtain the distance farthest between 2 in described isoline, contain described isoline at interior square with described distance for the length of side builds, obtain described foursquare area and using described foursquare area as described pollution effect area.

Step S12, utilizes described boundary layer parameters data, described profile data and described pollution effect area to set up the polynomial forecast model with predicting function.

Particularly, in the present embodiment, utilize described boundary layer parameters data, described profile data and described pollution effect area to set up the polynomial forecast model with predicting function to be specially: with described boundary layer parameters data and described profile data for training data, with described pollution effect area for dependent variable data, through obtaining multinomial model to the repeatedly training of described training data, more described multinomial model is optimized to the polynomial forecast model described in acquisition with predicting function according to square error and the goodness of fit.

Particularly, in the present embodiment, the polynomial forecast model described in predicting function is:

F (x ₁, x ₂, x ₃... x _m)=β ₀x ₁ ⁿ+ β ₁x ₂ ⁿ+ ...+β _n-1x _n ⁿ+ β _nx ₁ ^n-1x ₂... x _m; Wherein, x ₁, x ₂, x ₃... x _mbe respectively boundary layer parameters data and the principal component parameter of profile data after PCA analyzes (i.e. the training input data of polynomial forecast model); β ₀, β ₁... β _n-1, β _nbe respectively the coefficient of training input data in model; N is the number of model training input parameter; M is group item number.

According to square error and the goodness of fit described multinomial model is optimized and is specially: obtain the square error of model prediction area and and the goodness of fit after, utilize gradient descent method to obtain cost function value of coefficient in polynomial forecast model when minimum value.

Wherein, described model prediction area square error and be: wherein, S _ifor model is to the predicted value of area, S _rifor comprising the foursquare actual value of isoline, n is sample number, and i is i-th sample in sample; The described goodness of fit is: wherein, r ²for institute's established model is to the goodness of fit of problem matching, S _ifor model is to the predicted value of area, for the mean value of model prediction area, x _ithe training input data of the polynomial forecast model that the principal component parameter after PCA analyzes is formed for boundary layer parameters data and profile data; for x _imean value; Described cost function is: wherein, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model, r ²for institute's established model is to the goodness of fit of problem matching, S ⁽ⁱ⁾be the area actual value of i-th sample, f (X ⁽ⁱ⁾) be the predicted value of model to area of i-th sample; The formula that gradient descent method adopts is: wherein, α is learning rate, β _jfor the coefficient coefficient of polynomial forecast model training input data, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model.

Step S13, with real time meteorological data when there is atmospheric pollution for input, obtain prediction contaminated area according to described polynomial forecast model, the positional information of described prediction contaminated area and pollution source is imported in air pollution model as choosing the foundation evaluating and testing Polluted area in described air pollution model.

Below above-mentioned steps is further elaborated.

Be illustrated in figure 2 the schematic diagram of original AERMOD model, utilize original AERMOD model, the pollution concentration value of each predicted grid point is obtained according to history weather data, then interpolation is carried out to the concentration value of each lattice point, obtain the isoline of the emergent on-the-spot injury concentration paid close attention to, recycling algorithm draws the distance in isoline farthest between 2, with this distance for the length of side makes the square comprising this isoline, obtain its area value, as Fig. 3 signal.

Be specially: the meteorological pretreatment module AERMET process in AERMOD model of original ground weather data and sounding weather data, obtain the comprehensive weather data required for AERMOD nucleus module: boundary layer parameters data and profile data; In AERMAP module, set terrain elevation data file and choose evaluation and test region according to the requirement of master pattern and carry out stress and strain model, in conjunction with the parameter of on-the-spot actual scene and pollution source, be input in the lump in the middle of AERMOD nucleus module with the output of AERMET module after AERMAP terrain pre processing, calculate the concentration value of the pollutant of each net point ready-portioned in AERMAP; Then through the isoline of the interpolation acquisition certain concentration of interpolation algorithm, recycling algorithm draws the distance l in isoline farthest between 2, make for the length of side square 4 comprising this isoline with this distance al (0 < a < 0.5), obtain its area value S=al*al.

Because the size polluting the actual influence scope caused has reference role the most intuitively to being chosen at the evaluation and test region of it carrying out stress and strain model, consider again the promptness of emergent scene, therefore, utilize above-mentioned boundary layer parameters data and profile data, between specific isoline area corresponding with it, set up the forecast model with predicting function, be specially: the recurrence thought utilizing multinomial in machine learning, with boundary layer parameters data and profile data for training data, specific isoline area corresponding is with it dependent variable data, training through mass data obtains polynomial model, again by test data in conjunction with square error, the indexs such as the goodness of fit are carried out the test of algorithm model and are optimized rectification, obtain the good algorithm model of performance.

With the multinomial model of above-mentioned acquisition for forecast model, be applied in emergent scene and be: with the real-time weather data in scene for input, obtain prediction area, positional information etc. in conjunction with pollution source is converted to concrete longitude and latitude scope, import in the middle of AERMAP module, as choosing the foundation evaluating and testing region in AERMOD model with this scope.So, in the predicted application of realistic model, evaluation and test region by as much as possible close to the range of influence of actual contamination accident, decrease the calculating of unnecessary net point, improve forecasting efficiency.

As shown in the process flow diagram of the algorithm model of Fig. 4, the implementation process of the method mentioned in this patent is divided into three parts substantially:

1) according to weather data, pollution effect area is obtained: utilize original AERMOD model via AERMOD model and subsequent treatment, as shown in Figure 2, by AERMET module, pre-service is carried out to original weather data and obtain the comprehensive meteorological index and specific form that meet AERMOD module needs; In terrain pre processing module, some is the evaluation and test region comprising the actual Polluted area of accident, and another part is the landform altitude dem data region comprising evaluation and test region.Evaluation and test region is used for carrying out stress and strain model thereon, also the prediction carrying out contamination accident on this region is namely equivalent to, through the process of AERMAP, be input in the middle of AERMOD module, finally calculate the concentration value of the pollutant of each lattice point, then interpolation is carried out to the concentration value of each lattice point, obtain the isoline of the emergent on-the-spot injury concentration paid close attention to, recycling algorithm draws the distance in isoline farthest between 2, with this distance for the length of side makes the square comprising this isoline, obtain its area value, as Fig. 3 signal, and have: due to face phase emergency management and rescue, short-term prediction (namely the some time is carved into first 2 hours of this moment) is predicted as in our assumption method, hypothetical accident source 0 is true origin place on the way, and transverse axis is the direction of wind direction, the lower wind direction of pollution range in accident source of prediction, longitudinally considers by longitudinal diffusion theory.

Concrete steps are:

(1) long-term ground meteorological data parameter (.OQA file) and inspecting hole weather data file (.IQA file) is added up, by every day 8 time, 12 time, 16 time, 20 time, 24 time, 4 time weather data as the input of AERMET weather data pretreatment module in AERMOD model, the meteorological pretreatment module of the AERMET in AERMOD model as shown in Figure 1.

(2) output file of AERMET module is obtained: boundary layer parameters data file and profile data file.

(3) in AERMAP, set evaluation and test region (acceptance point) information of landform altitude information, pollution source dot information and stress and strain model, export the terrain data for AERMOD module and acceptance point data.

(4) above-mentioned boundary layer parameters data and profile data file, in conjunction with the actual parameter of pollution source 0, the output file of AERMAP terrain data pretreatment module as the input file of AERMOD module, through the calculating of nucleus module, obtain the concentration c of the contamination of each lattice point in AERMAP; Wherein in AERMOD module, the calculating of net point concentration value according to concrete scene, can calculate in conjunction with following algorithmic formula:

$C(X,Y,Z)=\frac{Q}{\stackrel{\&OverBar;}{u}}{p}_y(y,x)p_z(z,x)---\left(1\right)$

C _r(x,y,z)＝f.C(x _r,y _r,z _r)+(1-f).C(x _r,y _r,z _p) (2)

$C_d\{x_r,y_r,z_r\}=\frac{Q\&CenterDot;f_p}{\sqrt{2\mathrm{\&pi;u}}}\&CenterDot;F_y\&CenterDot;\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}\underset{m=0}{\overset{\&infin;}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&lambda;}}_j}{{\mathrm{\&sigma;}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-{\mathrm{\&psi;}}_{\mathrm{dj}}-{2\mathrm{mz}}_i)}^2}{{{\mathrm{z\&sigma;}}_{\mathrm{zj}}}^2}]+[\frac{{(z-{\mathrm{\&psi;}}_{\mathrm{dj}}-2{\mathrm{mz}}_i)}^2}{{{2\mathrm{\&sigma;}}_{\mathrm{zj}}}^2}\left]\right]$

$F_y=\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_y}\exp \left[\frac{{-y}^2}{{{2\mathrm{\&sigma;}}_y}^2}\right]---\left(3\right)$

$C_d\{x_r,y_r,z_r\}=\frac{Q\&CenterDot;f_p}{\sqrt{2\mathrm{\&pi;u}}}\&CenterDot;F_y\&CenterDot;\underset{j=1}{\overset{2}{\mathrm{\&Sigma;}}}\underset{m=1}{\overset{\&infin;}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&lambda;}}_j}{{\mathrm{\&sigma;}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-{\mathrm{\&psi;}}_{\mathrm{rj}}-{2\mathrm{mz}}_i)}^2}{{{\mathrm{z\&sigma;}}_{\mathrm{zj}}}^2}]+[\frac{{(z-{\mathrm{\&psi;}}_{\mathrm{rj}}-2{\mathrm{mz}}_i)}^2}{{{2\mathrm{\&sigma;}}_{\mathrm{zj}}}^2}\left]\right]---\left(4\right)$

$C_d\{x_r,y_r,z_r\}=\frac{Q\&CenterDot;\left({1-f}_p\right)}{\sqrt{2\mathrm{\&pi;u}}}\&CenterDot;F_y\&CenterDot;\mathrm{\&Sigma;}\underset{m=-\&infin;}{\overset{\&infin;}{\mathrm{\&Sigma;}}}\frac{{\mathrm{\&lambda;}}_j}{{\mathrm{\&sigma;}}_{\mathrm{zj}}}\left[\exp \right[\frac{{(z-h_{\mathrm{ep}}-{2\mathrm{mz}}_{\mathrm{ieff}})}^2}{{{\mathrm{z\&sigma;}}_{\mathrm{zp}}}^2}]+[\frac{{(z-h_{\mathrm{ep}}-2{\mathrm{mz}}_{\mathrm{ieff}})}^2}{{{2\mathrm{\&sigma;}}_{\mathrm{zp}}}^2}\left]\right]---\left(5\right)$

$C(x,y,z)=\frac{Q}{u}\&times;F_z\left(x_z{h}_p\right)\&times;F_y\left(y\right)$

$F_y=\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_y}\exp \left[\frac{{-y}^2}{{{2\mathrm{\&sigma;}}_y}^2}\right]$

$F_z=\frac{1}{\sqrt{2\mathrm{\&pi;}}{\mathrm{\&sigma;}}_z}\underset{n=-\&infin;}{\overset{\&infin;}{\mathrm{\&Sigma;}}}\exp \left[\frac{{(z-h_p+2{\mathrm{nh}}_a)}^2}{{{2\mathrm{\&sigma;}}_z}^2}\right]+\exp [\frac{{(z+h_p+{2\mathrm{nh}}_a)}^2}{{{2\mathrm{\&sigma;}}_z}^2}]---\left(6\right)$

Each meaning of parameters all can be learnt in technical background above.

(5) rational interpolation calculation is carried out to the concentration data of each lattice point of above-mentioned acquisition, draw the isoline 1,2,3 (c of the concentration (as sulphuric dioxide) that three are paid close attention to ₁< c ₂< c ₃), as three variable concentrations c in Fig. 2 ₁, c ₂, c ₃isoline signal.

(6) (concentration is c to calculate outermost layer isoline in above-mentioned three concentration isoline ₁) distance l between upper distance farthest 2, with the length of side of al (for coefficient, span is 0 < a < 0.5 to a) as square 4 in figure, make the square comprising this isoline, area is S=al*al.

2) with boundary layer data and profile data after PCA principal component analysis for training data (independent variable), the corresponding area comprising certain concentration isoline is dependent variable, use the recurrence thought of multinomial in machine learning, set up forecast model: because the size polluting the actual influence scope caused has reference role the most intuitively to being chosen at the evaluation and test region of it carrying out stress and strain model, consider again the promptness of emergent scene, therefore, utilize above-mentioned boundary layer parameters data and profile data, between specific isoline area corresponding with it, set up the forecast model with predicting function, be specially: the recurrence thought utilizing multinomial in machine learning, with boundary layer parameters data and profile data for training data, specific isoline area corresponding is with it dependent variable data, training through mass data obtains polynomial model, again by test data in conjunction with square error, the indexs such as the goodness of fit are carried out the test of algorithm model and are optimized rectification, obtain the good algorithm model of performance.Concrete steps are:

1. to AERMET export boundary layer data and profile data Z carry out PCA principal component analysis (PCA), choose a front m major component (expressing information amount (i.e. contribution rate of accumulative total) is greater than 85%) X, namely as the direct effect input training parameter of this patent method, area S is as dependent variable.

Z _i=(Z _i-μ _i)/δ _ii, μ _ifor Z _iaverage, δ _iifor Z _istandard deviation, then i-th major component is:

X _i=e _i' X, wherein e _ibe i-th proper vector;

The contribution rate of i-th major component is:

$G_i=\frac{{\mathrm{\&lambda;}}_i}{{\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2+{\mathrm{\&lambda;}}_3+...+{\mathrm{\&lambda;}}_m}$

2. in view of multinomial can describe comparatively complicated nonlinear problem, this patent chooses the polynomial expression with how far, by means of the thought of polynomial regression in machine learning, train the model of the multinomial that can describe this problem by the characterization factor X after process, shape as:

f(x ₁,x ₂,x ₃...x _m)＝β ₀x ₁ ⁿ+β ₁x ₂ ⁿ+...+β _n-1x _n ⁿ+β _nx ₁ ^n-1x ₂...x _n+...

3. with a large amount of training datas to above-mentioned model training, utilize gradient descent method to pass through to make the coefficient in the minimum Confirming model of cost function before variable, Confirming model;

Index:

Square error and: wherein S _ifor predicted value, S _rifor actual value;

The goodness of fit: $r^2=\frac{{\left[\mathrm{\&Sigma;}(x_i-\stackrel{\&OverBar;}{x})(S_i-\stackrel{\&OverBar;}{S})\right]}^2}{\mathrm{\&Sigma;}{(x_i-\stackrel{\&OverBar;}{x})}^2\mathrm{\&Sigma;}{(S_i-\stackrel{\&OverBar;}{S})}^2}$

Cost function: $J({\mathrm{\&beta;}}_0,{\mathrm{\&beta;}}_1\&CenterDot;\&CenterDot;\&CenterDot;)=\frac{1}{2r}\underset{i=0}{\overset{r}{\mathrm{\&Sigma;}}}{(f\left(X^{\left(i\right)}\right)-S^{\left(i\right)})}^2$

Gradient descent method: ${\mathrm{\&beta;}}_j={\mathrm{\&beta;}}_j-\mathrm{\&alpha;}\frac{\&PartialD;}{{\&PartialD;\mathrm{\&beta;}}_j}J({\mathrm{\&beta;}}_0,{\mathrm{\&beta;}}_1\&CenterDot;\&CenterDot;\&CenterDot;)$

4. by test data, above-mentioned model is tested again, carry out test of many times, according to selecting index appropriate model, cancelling noise point, then repeat above-mentioned training process, with the availability of verification model.

3) with scene obtain boundary layer parameters data and profile data for input, by the forecast model f (Z) of above-mentioned foundation, draw prediction area S, basis for selecting as evaluating and testing region in AERMAP: with the multinomial model of above-mentioned acquisition for forecast model f (Z), be applied in emergent scene and be: the boundary layer parameters weather data real-time with scene and profile data are input, obtain prediction area S, positional information etc. in conjunction with pollution source is converted to concrete longitude and latitude scope, import in the middle of AERMAP module with this scope, as choosing the foundation evaluating and testing region in AERMOD model.

Due to face phase emergency management and rescue, in our assumption method, be predicted as short-term prediction (namely the some time is carved into first 2 hours of this moment); Hypothetical accident source 0 is true origin place in fig. 3, and transverse axis is the direction of wind direction, and the pollution range of prediction considers the lower wind direction in accident source, longitudinally considers by longitudinal diffusion theory.This patent is in conjunction with model of atmospheric diffusion AERMOD, the size utilizing weather data and consequent pollution range is proposed, set up by weather data be driving factors, pollution range area is the forecast model of outcome variable, this model is utilized to dope corresponding contaminated area by new predicted data, in AERMAP module, the selection gist in the evaluation and test region of stress and strain model is carried out using this area, reduce the blindness in the past selected, improve the forecasting efficiency of model.Concrete step is played as described below:

1. add up long-term ground meteorological data parameter (.OQA file) and inspecting hole weather data file (.IQA file), by every day 8 time, 12 time, 16 time, 20 time, 24 time, 4 time weather data as the input of AERMET weather data pretreatment module in AERMOD model.

2. obtain the output file of AERMET module: boundary layer parameters data file and profile data file.

3. above-mentioned boundary layer parameters data and profile data file, in conjunction with the actual parameter of pollution source 0, the output file of AERMAP terrain data pretreatment module as the input file of AERMOD module, through the calculating of nucleus module, obtain the concentration of the contamination of each lattice point in AERMAP.

4. the concentration data of each lattice point of pair above-mentioned acquisition carries out rational interpolation calculation, draws the isoline 1,2,3 (c of the concentration (as sulphuric dioxide) that three are paid close attention to ₁< c ₂< c ₃).

5. calculate outermost layer isoline (c in above-mentioned three concentration isoline ₁) distance l between upper distance farthest 2, with al (0 < a < 0.5) for the length of side, make the square comprising this isoline, area is S=al*al.

6. the boundary layer data that export of couple AERMET and profile data X carry out PCA principal component analysis (PCA), choose a front m major component (expressing information amount (i.e. contribution rate of accumulative total) is greater than 85%) Z, namely as the direct effect input training parameter of this patent method, area S is as dependent variable.

Z _i=(Z _i-μ _i)/δ _ii, μ _ifor Z _iaverage, δ _iifor Z _istandard deviation, then i-th major component is:

X _i=e _i' X, wherein e _ibe i-th proper vector;

The contribution rate of i-th major component is:

$G_i=\frac{{\mathrm{\&lambda;}}_i}{{\mathrm{\&lambda;}}_1+{\mathrm{\&lambda;}}_2+{\mathrm{\&lambda;}}_3+...+{\mathrm{\&lambda;}}_m}$

7. in view of multinomial can describe comparatively complicated nonlinear problem, this patent chooses the polynomial expression with how far, by means of the thought of polynomial regression in machine learning, train the model of the multinomial that can describe this problem by the characterization factor X after process, shape as:

f(x ₁,x ₂,x ₃...x _m)＝β ₀x ₁ ⁿ+β ₁x ₂ ⁿ+...+β _n-1x _n ⁿ+β _nx ₁ ^n-1x ₂...x _n+...

8. with a large amount of training datas to above-mentioned model training, utilize gradient descent method to pass through to make the coefficient in the minimum Confirming model of cost function before variable, Confirming model.

Index:

Square error and: wherein S _ifor predicted value, S _rifor actual value.

The goodness of fit: $r^2=\frac{{\left[\mathrm{\&Sigma;}(x_i-\stackrel{\&OverBar;}{x})(S_i-\stackrel{\&OverBar;}{S})\right]}^2}{\mathrm{\&Sigma;}{(x_i-\stackrel{\&OverBar;}{x})}^2\mathrm{\&Sigma;}{(S_i-\stackrel{\&OverBar;}{S})}^2}$

Cost function: $J({\mathrm{\&beta;}}_0,{\mathrm{\&beta;}}_1\&CenterDot;\&CenterDot;\&CenterDot;)=\frac{1}{2r}\underset{i=0}{\overset{r}{\mathrm{\&Sigma;}}}{(f\left(X^{\left(i\right)}\right)-S^{\left(i\right)})}^2$

Gradient descent method: ${\mathrm{\&beta;}}_j={\mathrm{\&beta;}}_j-\mathrm{\&alpha;}\frac{\&PartialD;}{{\&PartialD;\mathrm{\&beta;}}_j}J({\mathrm{\&beta;}}_0,{\mathrm{\&beta;}}_1\&CenterDot;\&CenterDot;\&CenterDot;)$

9. by test data, above-mentioned model is tested again, carry out test of many times, according to selecting index appropriate model, cancelling noise point, then repeat above-mentioned training process, with the availability of verification model.

In sum, the present invention combines with model of atmospheric diffusion AERMO, proposes the method that optimum option carries out the evaluation and test region of stress and strain model, decreases the calculating consumption of unnecessary net point, improve forecasting efficiency; The method that the present invention proposes has predictability, reasonably can predict, reduce the blindness that evaluation and test region chosen by AERMOD model in the past of knowing clearly according to the true weather data before accident generation; The method that invention proposes has good applicability, can play reference role in good actual scene, the waste of manpower and materials in reduction accident in the middle of actual contamination accident to emergency management and rescue.So the present invention effectively overcomes various shortcoming of the prior art and tool high industrial utilization.

Above-described embodiment is illustrative principle of the present invention and effect thereof only, but not for limiting the present invention.Any person skilled in the art scholar all without prejudice under spirit of the present invention and category, can modify above-described embodiment or changes.Therefore, such as have in art usually know the knowledgeable do not depart from complete under disclosed spirit and technological thought all equivalence modify or change, must be contained by claim of the present invention.

Claims (8)

1. improve a method for air pollution model forecasting efficiency, be applied in the process by air pollution model evaluation and test Polluted area, it is characterized in that, described method comprises:

Pre-service is carried out to original weather data and obtains boundary layer parameters data and profile data, obtain the pollution concentration value of each predicted grid point in air pollution model according to described boundary layer parameters data, described profile data, pollution source parameter and terrain data, obtain the isoline of concern pollutant levels according to described pollution concentration value and obtain pollution effect area according to described isoline;

Described boundary layer parameters data, described profile data and described pollution effect area is utilized to set up the polynomial forecast model with predicting function;

With real time meteorological data when there is atmospheric pollution for input, obtain prediction contaminated area according to described polynomial forecast model, the positional information of described prediction contaminated area and pollution source is imported in air pollution model as choosing the foundation evaluating and testing Polluted area in described air pollution model.

2. the method for raising air pollution model forecasting efficiency according to claim 1, is characterized in that, by carrying out to the pollution concentration value of each predicted grid point described the isoline that interpolation arithmetic obtains described concern pollutant levels.

3. the method for raising air pollution model forecasting efficiency according to claim 2, it is characterized in that, described acquisition pollution effect area is specially: obtain the distance farthest between 2 in described isoline, contain described isoline at interior square with described distance for the length of side builds, obtain described foursquare area and using described foursquare area as described pollution effect area.

4. the method for the raising air pollution model forecasting efficiency according to claim 1,2 or 3, it is characterized in that, described isoline is at least two.

5. the method for raising air pollution model forecasting efficiency according to claim 1, is characterized in that, utilizes described boundary layer parameters data, described profile data and described pollution effect area to set up the polynomial forecast model with predicting function and is specially:

With described boundary layer parameters data and described profile data for training data, with described pollution effect area for dependent variable data, through obtaining multinomial model to the repeatedly training of described training data, more described multinomial model is optimized to the polynomial forecast model described in acquisition with predicting function according to square error and the goodness of fit.

6. the method for raising air pollution model forecasting efficiency according to claim 5, is characterized in that, described in there is predicting function polynomial forecast model be:

f(x ₁,x ₂,x ₃...x _m)＝β ₀x ₁ ⁿ+β ₁x ₂ ⁿ+...+β _n-1x _n ⁿ+β _nx ₁ ^n-1x ₂...x _m

Wherein, x ₁, x ₂, x ₃... x _mthe training input data of the polynomial forecast model that the principal component parameter being respectively boundary layer parameters data and profile data is formed; β ₀, β ₁... β _n-1, β _nbe respectively the coefficient of training input data in polynomial forecast model; N is the number of model training input parameter; M is the number of group item parameter.

7. the method for raising air pollution model forecasting efficiency according to claim 6, it is characterized in that, according to square error and the goodness of fit described multinomial model is optimized and is specially: obtain the square error of model prediction area and and the goodness of fit after, utilize gradient descent method to obtain cost function value of coefficient in polynomial forecast model when minimum value.

8. the method for raising air pollution model forecasting efficiency according to claim 7, is characterized in that,

The square error of described model prediction area and be: wherein, S _ifor model is to the predicted value of area, S _rifor comprising the foursquare actual value of isoline, n is sample number, and i is i-th sample in sample;

The described goodness of fit is: wherein, r ²for institute's established model is to the goodness of fit of problem matching, S _ifor model is to the predicted value of area, S is the mean value of model prediction area, x _ifor the training input data of the polynomial forecast model that the principal component parameter of boundary layer parameters data and profile data is formed; for x _imean value;

Described cost function is: wherein, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model, r ²for institute's established model is to the goodness of fit of problem matching, S ⁽ⁱ⁾be the area actual value of i-th sample, f (X ⁽ⁱ⁾) be the predicted value of model to area of i-th sample;

The formula that gradient descent method adopts is: wherein, α is learning rate, β _jfor the coefficient coefficient of polynomial forecast model training input data, β ₀, β ₁... be the coefficient of training input data in polynomial forecast model.