CN102799772B - Towards the sample optimization method of prediction of air quality - Google Patents

Abstract

The technical matters that the present invention solves be to provide a kind of forecast precision high, the sample optimization method towards prediction of air quality that adaptivity is good, comprise the following steps: (1) determines the correlation parameter that sample optimization screens, its specifically: the method determination meteorological factor weight matrix that (11) adopt pollutant levels and meteorological factor comprehensively to analyze; (12) orthogonal experiment method definite threshold matrix is adopted; (13) the method determination sample size of experimental check is adopted; (2) utilize the above-mentioned optimal screening that parameter carries out sample of determining, its specifically: (21) filter out each meteorological similarity and reach sample in threshold range; (22) filter out overall meteorological similarity and reach sample in threshold range; (23) filter out and predict the highest sample of day meteorological context similarity.

Description

Towards the sample optimization method of prediction of air quality

Technical field

The invention belongs to environmental forecasting field, particularly Air Quality Forecast field.

Background technology

From the nineties in 20th century, artificial neural network starts to be applied to Air Quality Forecast field.Artificial neural network plays the effect of data-driven in prediction of air quality process.To be calculated by neural network structure by input layer and draw output layer neuron.Neural network structure is adjusted by the learning process of historical data and is finally determined.

What require along with forecast precision improves constantly, and is also progressively developing the Application and improvement research of neural network.BP neural network is a kind of artificial neural network utilizing error backpropagation algorithm, by constantly repeating propagated forward and back-propagating process, until each training mode meets termination condition, then learn to terminate, thus utilized the network structure after study, draw according to prediction day input neuron information and predict the outcome.

From the aspect that model is set up, the factor of the e-learning that affects the nerves and generalization ability mainly contains three: network topology structure, learning algorithm and learning sample.To mainly concentrating on network structure and learning algorithm in the improvement of forecasting procedure, many scholars have done many research, improve the precision of prediction of network to a certain extent, achieve certain achievement.But after the improvement of network structure and learning algorithm acquires a certain degree, the raising of prediction of air quality classification accuracy is just limited to some extent.So outside improving network structure and learning algorithm, the selection of learning sample just becomes the final factor determining network mapping and generalization ability.Because the selection of learning sample, makes learning sample representative, can remove unnecessary interference, reaches the object of the excellent product of excellent kind.From physical angle, the principal element affecting atmospheric pollutant dispersion and conveying is meteorological condition, the sample optimization method set up based on meteorological similarity criterion becomes primary selection, because different meteorological factor has different impacts to pollutant levels, interaction between meteorological factor, the comprehensive weather environment formed also have Different Effects to pollutant, cause inconstant corresponding relation between meteorological Changing Pattern and pollutant levels.Polynary meteorologic factor moulds various meteorological spatial, and different meteorological spatial has different impacts to the transmission of pollutant and diffusion.The change of similarity to pollutant levels of meteorological spatial has regular reference.When carrying out prediction of air quality, if the meteorological spatial seeking suitable, the internal relation between multiple physical quantity and pollutant there has been reference.Select suitable sample set for major influence factors, make forecast work have more specific aim, improve network mapping and generalization ability, may ultimately reach the object improving prediction of air quality degree of accuracy.

Summary of the invention

The present invention solve technical matters be to provide a kind of forecast precision high, the good sample optimization method towards prediction of air quality with strong points of adaptivity.

For solving the problems of the technologies described above, the technical solution used in the present invention is: provide a kind of sample optimization method towards prediction of air quality, it is characterized in that, comprise the following steps:

(1) determine the correlation parameter that sample optimization screens, its specifically:

(11) adopt pollutant levels and meteorological factor analysis by synthesis method determination meteorological factor weight matrix, its specifically:

(111) meteorological factor weighted value under acquisition pollutant rise and fall sight;

(112) pollutant is obtained serious in meteorological factor weighted value under slight sight;

(113) related coefficient of wastewater mass concentration and meteorological factor is calculated;

(114) calculate the initial weight value of each meteorological factor, set up weight matrix;

(12) orthogonal experiment method definite threshold matrix is adopted;

(13) sample size is determined;

(2) filter out the sample of optimization, its specifically:

(21) filter out each meteorological similarity and reach sample in threshold range;

(22) filter out overall meteorological similarity and reach sample in threshold range;

(23) filter out and predict the highest sample of day meteorological context similarity.

The concrete steps of described step (111) are:

(1111) obtain sight meteorological factor representative data once, be pollutant concentration today and yesterday concentrations versus lifting value respectively at 0.05mg/m ³above and-0.05mg/m ³when following two interim each meteorological factor average and analyze the maximal value of this meteorological factor in the period, minimum value and mean value, i.e. upward period mean value M _iu, downward period mean value M _idand analyze the maximal value M of period _imax, minimum M _iminand mean value M _iadv, wherein i is meteorological factor label;

(1112) numerical value normalized, uses following formula process,

${M_{\mathrm{iu}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{iu}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{id}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{id}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{iadv}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{iadv}}}{M_{i\max }-M_{i\min }};$

(1113) meteorological factor change degree is calculated,

(1114) weighted value is calculated, wherein D _ifor meteorological factor change degree, n is meteorological factor species number.

The concrete steps of described step (112) are:

(1121) sight two times meteorological factor representative datas are obtained, the API value being pollutant respectively more than 100 and in less than 20 two periods each meteorological factor average and analyze the maximal value of this meteorological factor in the period, minimum value and mean value, namely seriously polluted period mean value M _iu, pollute mean value M in slight period _idand analyze the maximal value M of period _imax, minimum M _iminand mean value M _iadv, wherein i is meteorological factor label;

(1122) numerical value normalized, uses following formula process,

${M_{\mathrm{iu}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{iu}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{id}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{id}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{iadv}}}^{\′}=1-\frac{M_{i\max }-M_{\mathrm{iadv}}}{M_{i\max }-M_{i\min }};$

(1123) meteorological factor change degree is calculated,

(1124) weighted value is calculated, wherein D _ifor meteorological factor change degree, n is meteorological factor species number.

Described in described step (113), the related coefficient of wastewater mass concentration and meteorological factor is:

in formula, i is meteorological factor label; N meteorological factor species number; x _i

For wastewater mass concentration value; for wastewater mass concentration mean value; y _ifor meteorological factor value; for meteorological factor mean value.

Described in described step (114), the initial weight value of each meteorological factor is:

$R_i=\frac{1}{2}(w_{1i}+w_{2i})\·r_i$

In formula, r _ifor related coefficient; w ¹ⁱand w ²ⁱbe respectively sight one and sight two times calculating gained weighted values.The concrete steps of step (12) are:

(121) determine that experimental index is the absolute error value of model prediction result;

(122) select factor, determine level, the experimental factor that need investigate is determined according to the result of (11) step weight matrix, namely choose in this patent to the large meteorologic factor of pollutant levels weighing factor value as experimental factor, and draft the level of each factor;

(123) design orthogonal arrage gauge outfit, experimental factor is filled into respectively the row of orthogonal arrage;

(124) each level numeral is inserted the row of orthogonal arrage;

(125) calculating K _jmvalue, wherein K _jmtest index corresponding to jth row factor m level and;

(126) the mean value k of each factor same level is calculated _jm, k _jmfor K _jmmean value

k _jm＝K _jm/4；

(127) the extreme difference R of each factor is calculated _j

R _j＝max(k _j)-min(k _j)

R _jrepresent the amplitude of jth row factor test index change in its span;

(128) optimal combination, according to the mean value k of each level of each factor _jmdetermine excellent level.

The determination of described step (13) sample size adopts the method for experimental check, and adopt multiple sample size to test respectively for often kind of pollutant, the sample size that Selection Model prediction effect is best, namely precision of prediction is the highest.

The Filtering system of described step (2) Screening Samples is similar for criterion with meteorology.

The concrete steps of described step (2) are:

(21) filter out each meteorological similarity and all reach the sample of specifying in threshold range, concrete grammar is:

Sample must meet following formula:

Δ y _j≤ y _jset, wherein, Δ y _j=| y _{j is pre-}-y _{j sample}|

In formula, y _{j is pre-}for predicting the meteorological factor value of day;

Y _{j sample}for the meteorological factor value of sample;

Δ y _jfor the meteorological similarity between sample and prediction day each meteorological factor;

J is meteorological factor label;

Y _jsetfor the threshold value that each meteorological factor screens, composition initial threshold matrix Y, the threshold value in this matrix can according to the dynamic change of sample requirement amount;

(22) filter out overall meteorological similarity and reach sample in certain threshold range, the sample filtered out must meet following formula:

S≤S _set, wherein, $S=\underset{j\≤\mathrm{Mnum}}{\mathrm{\Σ}}(w_j\·\mathrm{\Δ}{y}_j)$

In formula, S is overall meteorological similarity;

S _setfor the threshold value of overall meteorological similarity screening;

W _jfor the weight of each meteorological factor of sample, composition weight matrix W, reflects that this meteorologic factor is to pollutant levels influence degree;

M _numfor the number of meteorological factor;

(23) filter out and predict the highest n bar sample of day meteorological context similarity.The sample filtered out must meet following formula:

Q _num≤n

In formula, Q is with the ascending order sample row of overall meteorological similarity S sequence;

Q _numfor the sequence number of sample in the sample row after sequence;

N is the sample size of demand.

Compared with the existing Urban Air Pollution Methods based on neural network, three layers of sample optimization screening technique of meteorological similarity criterion make the prediction of air quality of artificial neural network have the following advantages:

1, sample optimization method to often kind of pollutant all pointed and practicality.The concentration change impact of identical meteorological spatial field on different pollutant is different.Often kind of pollutant all has the feature meteorological factor affecting its concentration change, and they are to pollutant effects degree varies.By determine to affect often kind of pollutant feature meteorological factor and to pollutant effects degree, the sample of optimization will have more specific aim.

2, sample optimization method is pointed to the change of prediction day pollutant levels.There is larger difference in the meteorological spatial field of different times, although period is close, due to the impact by cold air, typhoon or other special weathers, the meteorological spatial field on close date also may not be identical.To predict that same day day and background meteorology thereof are as optimizing criterion, by searching relative meteorological condition in historical data base, choosing the most close meteorological spatial field, setting up forecasting model with this.For prediction day pollutant situation of change, specific aim will be had more.

Accompanying drawing explanation

Fig. 1 is three layers of sample optimization screening process;

Fig. 2 is ground floor screening sample flow process;

Fig. 3 is second layer screening sample flow process;

Fig. 4 is third layer screening sample flow process;

Fig. 5 is orthogonal experiments analysis process;

Fig. 6 is weight matrix defining method flow process.

Embodiment

Come that the present invention will be further described in conjunction with the accompanying drawings and embodiments.

The invention provides a kind of sample optimization method towards prediction of air quality, it is characterized in that, comprise the following steps:

(1) determine the correlation parameter that sample optimization screens, its specifically:

(11) adopt pollutant levels and meteorological factor analysis by synthesis method determination meteorological factor weight matrix, its specifically:

(111) meteorological factor weighted value under acquisition pollutant rise and fall sight;

(112) pollutant is obtained serious in meteorological factor weighted value under slight sight;

(113) related coefficient of wastewater mass concentration and meteorological factor is calculated;

(114) calculate the initial weight value of each meteorological factor, set up weight matrix;

(12) orthogonal experiment method definite threshold matrix is adopted;

(13) sample size is determined;

(2) Screening Samples, obtains the sample optimized.

Example below in conjunction with reality is next, and present invention is described: prepare data: environment automatic monitoring station, somewhere PM10 mass concentration, meteorologic factor (temperature, air pressure, wind speed, wind direction, solar radiation, rainfall amount, relative humidity).

(1) sample optimization screening correlation parameter is determined

(11): the determination of weight matrix

(111): pollutant rise and fall sight meteorological factor weighted value

(1111): shown in meteorological factor typical value is calculated as follows:

Table 1 meteorological factor typical value

(1112): numerical value normalization calculates

(1113): meteorological factor change degree calculates

(1114): weighted value result of calculation

Meteorological factor weighted value w under this sight is obtained by above four steps _1i:

Table 2 is polluting the meteorological factor weighted value risen and in decline sight

(112): the analysis in seriously polluted and slight period

(1121): shown in meteorological factor typical value is calculated as follows:

Table 3 meteorological factor typical value

(1122): numerical value normalization calculates

(1123): meteorological factor change degree calculates

(1124): weighted value result of calculation

Meteorological factor weighted value w under this sight is obtained by above four steps _2i:

The meteorological factor weighted value of table 4 in seriously polluted and slight scenario analysis

(113): Calculation of correlation factor result

Table 5 pollutant levels and meteorological factor related coefficient

(114): determine weight matrix

Carry out the screening of influence factor and the correction of weight through overtesting, finally set up the weight matrix of sample optimization.

Table 6 sample optimization weight matrix

(12): the determination of threshold matrix

(121): the best of clear and definite experiment purpose-find out sample optimization combines with it; Determine experimental index--with forecast model absolute error AE as test index.

(122): select factor, level is determined.Determining four experimental factors according to the result of weight matrix: wind, relative humidity, rainfall grade, air pressure, for investigating the nonlinear relationship of factor and PM10 error prediction model, respectively selecting three levels.As shown in table 7.

Table 7 factor level table

(123): selected orthogonal arrage, design gauge outfit.The orthogonal arrage selected is L9 (34), and four factors are arranged in A, B, C, D tetra-respectively and arrange.

(124): level is translated, establishment experimental program.Do 9 tests altogether, testing program is as shown in table 8.

(125): test by scheme, record test index end value, in table 8 " model AE " row, analyzes experimental result.Process and result as shown in table 8, calculating K _jmvalue.

As, the value of K I first row to represent under first row factor (A wind) I level corresponding test index and, i.e. 0.0127+0.0119+0.0128=0.0374, the value that K III the 2nd arranges represent experimental index corresponding under secondary series factor (B humidity) II level and, i.e. 0.0128+0.0126+0.0115=0.0369.The like.

(126): the mean value k calculating each factor same level _jm.

(127): the extreme difference R calculating each factor _j, according to the primary and secondary impact order of extreme difference decision factor.

(128): select excellent combination.According to the mean value k of each level of each factor _jmdetermine excellent level.As for Elements C (rainfall grade), there is k II <k I <k III, therefore optimal level is II.The like, the optimum combination finally determined is A III B III C II D I.

Table 8 is for the sample optimization Threshold selection testing program of PM10 forecast model and interpretation of result

Wherein proxima luce (prox. luc) threshold value increases to some extent on the basis of experimental result, and except rainfall grade increases by 1, all the other parameters all increase by 2.Setting population sample threshold value is 5.Set up the threshold matrix of sample optimization thus, as shown in table 9.

Table 9 sample optimization threshold matrix

(13): the determination of sample size

Test with different sample size, obtain precision when sample size is 330 the highest, therefore determine that the sample size for training PM10 neural network is 330.

(2) sample optimization screening

(21): the method according to step 21 carries out ground floor optimal screening to sample, namely filter out each meteorologic factor (today wind, rainfall today, rainfall yesterday, yesterday air pressure, yesterday humidity) similarity all reaches the sample of (see table 9) in given threshold range.

(22): the method according to step 22 carries out second layer optimal screening to the initial sample set obtained in previous step, the sample that overall meteorological similarity reaches (<=5) within the scope of defined threshold is namely filtered out.

(23): to the double optimization sample set obtained in previous step carry out with s be order standard ascending order arrangement, obtain sample sequence Q, if sample number is less than 330, then all retain, if more than 330, then filter out and predict the highest front 330 samples of day meteorological context similarity.

So far, whole work of screening sample are completed.

Claims (9)

1., towards a sample optimization method for prediction of air quality, it is characterized in that, comprise the following steps:

(1) determine the correlation parameter that sample optimization screens, its specifically:

(11) adopt pollutant levels and meteorological factor analysis by synthesis method, determine meteorological factor weight matrix, its specifically:

(111) meteorological factor weighted value under acquisition pollutant rise and fall sight;

(112) pollutant is obtained serious in meteorological factor weighted value under slight sight;

(113) related coefficient of wastewater mass concentration and meteorological factor is calculated;

(114) calculate the initial weight value of each meteorological factor, set up weight matrix;

(12) orthogonal experiment method definite threshold matrix is adopted;

The concrete steps of step (12) are:

(121) determine that experimental index is the absolute error value of model prediction result;

(122) select factor, determine level, determine the experimental factor that need investigate according to the result of step (11) weight matrix, namely choose to the large meteorologic factor of pollutant levels weighing factor value as experimental factor, and draft the level of each factor;

(123) design orthogonal arrage gauge outfit, experimental factor is filled into respectively the row of orthogonal arrage;

(124) each level numeral is inserted the row of orthogonal arrage;

(125) calculating K _jmvalue, wherein K _jmtest index corresponding to jth row factor m level and;

(126) the mean value k of each factor same level is calculated _jm, k _jmfor K _jmmean value,

k _jm＝K _jm/4；

(127) the extreme difference R of each factor is calculated _j,

R _j＝max(k _j)-min(k _j)，

R _jrepresent the amplitude of jth row factor test index change in its span;

(128) optimal combination, according to the mean value k of each level of each factor _jmdetermine excellent level;

(13) the method determination sample size of experimental check is adopted;

(2) filter out the sample of optimization, it is specially:

(21) filter out each meteorological similarity and reach sample in threshold range;

(22) filter out overall meteorological similarity and reach sample in threshold range;

(23) filter out and predict the highest sample of day meteorological context similarity.

2. the sample optimization method towards prediction of air quality according to claim 1, is characterized in that, the concrete steps of described step (111) are:

(1111) obtain sight meteorological factor representative data once, be pollutant concentration today and yesterday concentrations versus lifting value respectively at 0.05mg/m ³above and-0.05mg/m ³when following two interim each meteorological factor average and analyze the maximal value of this meteorological factor in the period, minimum value and mean value, i.e. upward period mean value M _iu, downward period mean value M _idand analyze the maximal value M of period _imax, minimum M _iminand mean value M _iadv, wherein i is meteorological factor label;

(1112) numerical value normalized, uses following formula process,

${M_{\mathrm{iu}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{iu}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{id}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{id}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{iadv}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{iadv}}}{M_{i\max }-M_{i\min }};$

(1113) meteorological factor change degree is calculated,

(1114) weighted value is calculated, wherein D _ifor meteorological factor change degree, n is meteorological factor species number.

3. the sample optimization method towards prediction of air quality according to claim 2, is characterized in that, the concrete steps of described step (112) are:

(1121) sight two times meteorological factor representative datas are obtained, the API value being pollutant respectively more than 100 and in less than 20 two periods each meteorological factor average and analyze the maximal value of this meteorological factor in the period, minimum value and mean value, namely seriously polluted period mean value M _iu, pollute mean value M in slight period _idand analyze the maximal value M of period _imax, minimum M _iminand mean value M _iadv, wherein i is meteorological factor label;

(1122) numerical value normalized, uses following formula process,

${M_{\mathrm{iu}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{iu}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{id}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{id}}}{M_{i\max }-M_{i\min }},{M_{\mathrm{iadv}}}^{\&prime;}=1-\frac{M_{i\max }-M_{\mathrm{iadv}}}{M_{i\max }-M_{i\min }};$

(1123) meteorological factor change degree is calculated,

(1124) weighted value is calculated, wherein D _ifor meteorological factor change degree, n is meteorological factor species number.

4. the sample optimization method towards prediction of air quality according to claim 1, is characterized in that, described in described step (113), the related coefficient of wastewater mass concentration and meteorological factor is:

in formula, i is meteorological factor label; N meteorological factor species number; x _ifor wastewater mass concentration value; for wastewater mass concentration mean value; y _ifor meteorological factor value; for meteorological factor mean value.

5. the sample optimization method towards prediction of air quality according to claim 3, is characterized in that, described in described step (114), the initial weight value of each meteorological factor is:

$R_i=\frac{1}{2}(w_{1i}+w_{2i})\&CenterDot;r_i$

In formula, r _ifor related coefficient; w _1iand w _2ibe respectively sight one and sight two times calculating gained weighted values.

6. the sample optimization method towards prediction of air quality according to claim 1, is characterized in that, in step (123), factor number should be not more than the columns of orthogonal arrage.

7. the sample optimization method towards prediction of air quality according to claim 1, it is characterized in that, the determination of described step (13) sample size adopts the method for experimental check, adopt multiple sample size to test respectively for often kind of pollutant, choose the sample size that precision of prediction is the highest.

8. the sample optimization method towards prediction of air quality according to claim 1, is characterized in that, the Filtering system of described step (2) Screening Samples is similar for criterion with meteorology, and the concrete steps of described step (2) are:

(21) filter out each meteorological similarity and reach sample in threshold range, concrete grammar is:

Sample must meet following formula:

Δ y _j≤ y _jset, wherein, Δ y _j=| y _{j is pre-}-y _{j sample}|,

In formula, y _{j is pre-}for predicting the meteorological factor value of day;

Y _{j sample}for the meteorological factor value of sample;

Δ y _jfor the meteorological similarity between sample and prediction day each meteorological factor;

J is meteorological factor label;

Y _jsetfor the threshold value that each meteorological factor screens, composition initial threshold matrix Y, the threshold value in this matrix can according to the dynamic change of sample requirement amount;

(22) filter out overall meteorological similarity and reach sample in certain threshold range, the sample filtered out must meet following formula:

S≤S _set, wherein, $S=\underset{j\&le;\mathrm{Mnum}}{\mathrm{\&Sigma;}}(w_j\&CenterDot;\mathrm{\&Delta;}{y}_j),$

In formula, S is overall meteorological similarity;

S _setfor the threshold value of overall meteorological similarity screening;

W _jfor the weight of each meteorological factor of sample, composition weight matrix W, reflects that this meteorologic factor is to pollutant levels influence degree;

M _numfor the number of meteorological factor;

(23) filter out and predict the highest n bar sample of day meteorological context similarity, the sample filtered out must meet following formula:

Q _num≤n，

In formula, Q is with the ascending order sample row of overall meteorological similarity S sequence;

Q _numfor the sequence number of sample in the sample row after sequence;

N is the sample size of demand.

9. the sample optimization method towards prediction of air quality according to any one of claim 1-8, is characterized in that, described meteorological factor is temperature, air pressure, wind speed, wind direction, solar radiation, rainfall amount or relative humidity.