CN110046771A - A kind of PM2.5 concentration prediction method and apparatus - Google Patents

Abstract

The present invention relates to a kind of PM2.5 concentration prediction method and apparatus, belong to the electric powder prediction of PM2.5 concentration value.Wherein method is the following steps are included: obtain PM2.5 concentration history data, meteorological historical data, Determination of Aerosol Optical historical data；Obtain two-dimensional position information, elevation information and the sampling time information of observation point；The parameter of 4D-GTWR model is determined according to above-mentioned data, so that it is determined that 4D-GTWR model, 4D-GTWR model are as follows:According to the 4D-GTWR model of the meteorological data of prediction, Determination of Aerosol Optical data and determination, the concentration of PM2.5 is predicted.4D-GTWR model in this method provides effective method in terms of predicting PM2.5 concentration, so that the prediction of PM2.5 concentration is more accurate, while also solving the non-stationary of four-dimensional spacetime, also demonstrating elevation information has great influence to the variation of PM2.5 concentration.

Description

A PM2.5 concentration prediction method and device

technical field

The invention relates to a PM2.5 concentration prediction method and device, belonging to the technical field of PM2.5 concentration value prediction.

Background technique

Fine particulate matter (PM2.5) not only poses a serious threat to public health, but also seriously affects urban traffic and citizens' daily lives. Therefore, PM2.5, as an air pollutant that directly affects human life and health, has attracted many scholars' attention. The extensive attention of PM2.5 related inversion research work to realize the prediction of PM2.5 concentration.

There are many methods for predicting PM2.5 in the prior art. For example, the Chinese patent application document with application publication number CN106056210A discloses a method for predicting PM2.5 concentration value based on a hybrid neural network. The method adopts PM2.5 concentration The historical data of the value, the historical data of the relevant indicators, the historical meteorological data and the analysis data of the PM2.5 components are used to simulate the variation law of the local PM2.5 concentration value through the neural network, so as to realize the prediction of the PM2.5 concentration value.

Another example: Zhao Yangyang et al. combined collaborative training with the spatiotemporal geographic weighted regression model (GTWR), and proposed a collaborative spatiotemporal geographic weighted regression PM2.5 concentration estimation method. The source is "Science of Surveying and Mapping", 2016, 41(12): 172-178, this method takes the Beijing-Tianjin-Hebei PM2.5 concentration from March to July 2015 as an example, adopts the Euclidean spatial distance metric, and conducts a comparative analysis experiment through GTWR of different kernel functions. The results show that the number of samples in space and time is insufficient. , the method can improve the estimation accuracy of PM2.5 concentration. However, the prediction accuracy of the PM2.5 concentration prediction method in the prior art still needs to be improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a PM2.5 concentration prediction method to solve the problem of low accuracy of the existing prediction method; and also to provide a PM2.5 concentration prediction device to solve the accuracy of the existing prediction device. low sex issue.

In order to achieve the above object, the present invention proposes a PM2.5 concentration prediction method, comprising the following steps:

Obtain PM2.5 concentration historical data, meteorological historical data, and atmospheric aerosol optical depth historical data;

Obtain two-dimensional position information, elevation information and sampling time information of observation points;

According to the above data, the parameters of the 4D-GTWR model are determined to determine the 4D-GTWR model. The 4D-GTWR model is:

Wherein, x _ik is the k-th independent variable of observation point i, there are p independent variables in total, and the independent variable is meteorological data; y _i is the dependent variable of observation point i, and the dependent variable is PM2.5 concentration data; The parameters of the 4D-GTWR model are: β _ik (u _i ,vi ,z _i ,t _i ), β _i0 (u _i ,vi ,z _i ,t _{i )} and ξ _i , β _ik ( _{u i ,} v _i , z _i , t _i ) is the independent variable regression coefficient of the k-th independent variable of observation point i, which is related to the spatial position of observation point i; β _i0 (u _i ,v _i ,z _i ,t _i ) is the observation point i The independent variable regression coefficient constant term of point i; ξ _i is the random error of observation point i; (u _i ,vi ,z _i ,t _i ) is the four-dimensional space coordinate of observation point _i , where (u _i ,vi _i ) represents the two-dimensional space coordinate, (z _i ) represents the elevation space coordinate, (t _i ) represents the time coordinate; n is the number of observation points;

Based on the predicted meteorological data, atmospheric aerosol optical depth data and the determined 4D-GTWR model, the PM2.5 concentration was predicted.

In addition, the present invention also provides a PM2.5 concentration prediction device, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, the processor implements when the computer program is executed The above-mentioned PM2.5 concentration prediction method.

The beneficial effects are: the elevation information is introduced into the space-time geography weighted regression model, and a four-dimensional space-time geography weighted regression model (4D-GTWR) is proposed, which adopts the four-dimensional space Euclidean distance measure, which can better reflect the four-dimensional space-time distribution characteristics and provide more The real spatial distance can effectively analyze the actual situation, and the 4D-GTWR model provides an effective method in predicting PM2. Stationarity also verifies that the elevation information has an important influence on the change of PM2.5 concentration.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the method for calculating the regression coefficient of the independent variable includes: by establishing the objective function of the observation point i, the least squares estimation of the regression coefficient of the independent variable is obtained, and the objective function is:

Among them, w _ij ^ST is the kernel function between observation point i and observation point j;

least squares estimates of the regression coefficients of the independent variables for:

Among them, W(u _i ,v _i ,z _i ,t _i ) is a four-dimensional space-time weight matrix, which is a matrix composed of w _ij ^ST , X is the independent variable matrix, and Y is the observation value matrix of the dependent variable.

The beneficial effects are: by establishing an objective function to solve the regression coefficient of the independent variable, the method is simple and accurate.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the kernel function is a Gaussian type kernel function, and the formula is:

Among them, h ^4D-ST is the four-dimensional space-time bandwidth, and d _ij ^ST is the four-dimensional space-time distance between observation point i and observation point j.

The beneficial effect is that the four-dimensional space-time weight matrix can be obtained more accurately through the Gaussian kernel function.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the elevation information is digital elevation information, and the calculation formula of the four-dimensional space-time distance between the observation point i and the observation point j is:

in, is the two-dimensional spatial distance between observation point i and observation point j; is the digital elevation space distance between observation point i and observation point j; is the time distance between observation point i and observation point j; λ, δ and μ are scale adjustment factors, which are used to balance the scale differences of different four-dimensional space-time distances.

The beneficial effects are: the method for calculating the four-dimensional space-time distance is simple and accurate.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the four-dimensional space-time bandwidth is obtained according to the two-dimensional space bandwidth h ^S2d , the digital elevation space bandwidth h ^SDEM and the time bandwidth h ^T .

The beneficial effects are: solving the two-dimensional space bandwidth, the digital elevation space bandwidth and the time bandwidth respectively, and then obtaining the four-dimensional space-time bandwidth according to the two-dimensional space bandwidth, the digital elevation space bandwidth, the time bandwidth and the four-dimensional space-time bandwidth, so that the obtained four-dimensional space-time bandwidth is more accurate. .

Further, in the above-mentioned PM2.5 concentration prediction method and device, two-dimensional space bandwidth, digital elevation space bandwidth and time bandwidth are calculated according to the Akaike Information Content Criterion, and the Akaike Information Content Criterion AICc is:

Among them, σ ² is the unbiased estimate of the random error variance in the 4D-GTWR model, S is the hat matrix of the 4D-GTWR model, X ₁ , X ₂ , ..., X _n are the values of observation points 1, 2, ..., n, respectively A row vector of independent variables.

The beneficial effects are: more accurate bandwidth data can be obtained through the Akaike information quantity criterion, and the accuracy of PM2.5 prediction can be further improved.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the calculation formula of the unbiased estimate σ ² of the random error variance in the 4D-GTWR model is:

σ ² =RSS _4D-GTWR /(n-2tr(S)+tr(S ^T S));

Among them, RSS _4D-GTWR is the residual sum of squares of the 4D-GTWR model.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the calculation formula of the residual squared sum RSS _4D-GTWR of the 4D-GTWR model is:

RSS _4D-GTWR = Y ^T (IS) ^T (IS) Y;

where I is the identity matrix.

Further, in the above-mentioned PM2.5 concentration prediction method and device, the meteorological data includes at least air pressure, air temperature, relative humidity, rainfall, wind speed and wind direction.

The beneficial effects are: the above meteorological data are all related to the concentration of PM2.5, and the concentration of PM2.5 can be more comprehensively predicted through the above meteorological data.

Description of drawings

1 is a schematic diagram of the four-dimensional space-time distance between observation point i and observation point j of the present invention;

Fig. 2 is the relation diagram of PM2.5 estimated value and observed value using GTWR model in the prior art;

FIG. 3 is a graph showing the relationship between the PM2.5 estimated value and the observed value using the 4D-GTWR model in the present invention.

Detailed ways

Example of PM2.5 concentration prediction method:

The PM2.5 concentration prediction method proposed in this embodiment includes the following steps:

1) Obtain historical data of PM2.5 concentration, historical meteorological data, and historical data of atmospheric aerosol optical depth.

Different meteorology and different atmospheric aerosol optical depth (AOD data) will have different effects on PM2.5 concentration.

In this embodiment, the meteorological factors included in the meteorological historical data include air pressure (h Pa), air temperature (° C.), relative humidity (%), rainfall (mm), wind speed (km/h) and wind direction (°). However, wind speed, wind direction and relative humidity have a greater impact on PM2.5 concentration, so only wind speed, wind direction and relative humidity are acceptable in historical meteorological data.

The historical data of PM2.5 concentration is detected by the air quality pollution automatic monitoring station (hereinafter referred to as the monitoring station), which is the monitoring data of the average concentration of PM2.5 per hour, and the historical meteorological data is detected by the weather station, which is the average value per hour data. These data can be downloaded from the website of the National Meteorological Information Center of China. Due to the geographical differences between the meteorological sites and the PM2.5 monitoring sites, the values of the meteorological factors of the PM2.5 monitoring sites are obtained from the data of the surrounding meteorological sites. Gold interpolation comes.

The AOD data is retrieved by the Moderate Resolution Imaging Spectrometer (MODIS), and the MYD04_3K data is geometrically corrected by the IDL program design, converted into the WGS-84 geographic coordinate system, and extracted by Arcpy to match the PM2.5 detection site in space and time AOD data. MODIS is a commonly used sensor for retrieving the optical depth of atmospheric aerosols, with a scanning width of 2330 km, which can obtain global observation data at least once a day. MODIS has 36 spectral channels, the spectral range is 0.4–14 μm, and the spatial resolution can be 250m, 500m, 1000m, and can be used to obtain AOD, water vapor, surface temperature, ocean and other data. The AOD data in the MODIS data is provided by NASA LAADS for data download.

2) Obtain the two-dimensional position information, digital elevation information (ie DEM information) and sampling time information of the observation point.

The two-dimensional position information is the position information of the observation point, and the sampling time is from 2017.12 to 2018.2. In this embodiment, digital elevation information (ie DEM information) is used as the height information of the three-dimensional space, and the DEM data (ie DEM information) comes from the computer of the Chinese Academy of Sciences. The network information center geospatial data cloud platform, the spatial resolution is 30m, and the DEM data is the data in 2009. As other implementation manners, other elevation information may also be used, which is not limited in the present invention.

3) Determine the parameters of the 4D-GTWR model according to the above data, thereby determining the 4D-GTWR model.

The 4D-GTWR model is:

Among them, x _ik is the k-th independent variable of observation point i, there are p independent variables in total, and the independent variable is meteorological data; y _i is the dependent variable of observation point i, and the dependent variable is PM2.5 concentration data; 4D-GTWR model The parameters are: β _ik (u _i ,vi ,z _i ,t _i ), β _i0 (u _i ,vi ,z _i ,t _i ) and ξ _i , β _ik ( _{u i} ,vi _, z _{i )} ,t _i ) is the independent variable regression _coefficient of the _k - _th independent variable of observation point _i , which is related to the spatial position of observation point _i ; Variable regression coefficient constant term; ξ _i is the random error of observation point i; (u _i ,vi ,z _i ,t _i ) is the four-dimensional space coordinate of observation point _i , where ( _{u i} ,vi ) represents two-dimensional Space coordinates, (z _i ) represents digital elevation spatial coordinates, (t _i ) represents time coordinates; n is the number of observation points.

The solution to this model is mainly to solve the independent variable regression coefficient β _ik (u _i ,vi ,z _i ,t _i ), ξ _i is the random error of observation point _i , obeying ξ _i -N(0,ω ² ) is independent and identically distributed, and Cov(ξ _i , ξ _j )=0 (i≠j) (that is, random variables ξ _i and ξ _j have the same probability distribution and are independent of each other).

The least squares estimate of the independent variable regression coefficient for observation point i is is a matrix composed of β _ik (u _i ,v _i ,z _i ,t _i ), and β _i0 (u _i ,v _i ,z _i ,t _i ) is The value of the first column in the matrix is used to estimate the 4D-GTWR model according to the weighted least squares criterion, and an objective function is established for each observation point. The objective function of observation point i is as follows:

Among them, w _ij ^ST is the kernel function (ie the weight) between the observation point i and the observation point j, which is related to the four-dimensional space-time distance.

Least Squares Estimation of Regression Coefficients of Independent Variables can be expressed as:

Among them, W(u _i ,v _i ,z _i ,t _i ) represents the four-dimensional space-time weight matrix, which is a matrix composed of w _ij ^ST , X is the independent variable matrix, and 1 in the X matrix is the independent variable x corresponding to β _{i0 i0} , generally takes the value of 1, Y is the observed value matrix (here is the observed actual PM2.5 value), the estimated value of the dependent variable of the observation point i can be obtained through the above calculation for:

Among them, X _i represents the vector of the i-th row of the matrix X (that is, the row vector composed of the independent variables of the observation point i). Therefore, the dependent variable regression vector at each observation point (ie, the predicted outcome) for:

where S is the hat matrix of the 4D-GTWR model.

In this embodiment, the w _ij ^ST kernel function is a Gaussian type kernel function. As another implementation manner, the w _ij ^ST kernel function may also be a Bi-square type kernel function, and the present invention does not limit the type of the kernel function.

The Gaussian kernel function formula is:

Among them, h ^4D-ST is the four-dimensional space-time bandwidth, and d _ij ^ST is the four-dimensional space-time distance between observation point i and observation point j.

Considering the heterogeneity of four-dimensional space in three-dimensional space and one-dimensional time, the present invention proposes an analysis and calculation method for four-dimensional space-time distance.

Considering that DEM mainly has a certain impact on the spatial dimension, based on the research on the temporal and spatial variation of PM2.5 in this paper, the pollution process of PM2.5 and other pollutants has a four-dimensional spatial-temporal regional variation, that is, there is a four-dimensional spatial-temporal non-stationary change. Therefore, for any number of monitoring points, the three-dimensional space coordinates between them are different. Considering that two- and three-dimensional spaces may have different scale effects, the 4D-GTWR model introduces to show the difference in scale between them. Therefore, according to the Euclidean space distance, the three-dimensional space distance d _ij ^s between the observation point i and the observation point j can be expressed as follows:

in, Various operators can be represented. On this basis, combining the spatiotemporal distance construction method, the four-dimensional spatiotemporal distance is expressed as follows:

scale effect symbol and Usually, addition is used, that is, the linear combination of the four-dimensional space-time distance can be obtained. Under the 4D-GTWR model, as shown in Figure 1, the observation point i (that is, the regression point in Figure 1) and the observation point j (that is, the The four-dimensional space-time distance d _ij ^ST between adjacent points) is expressed as follows:

in, is the two-dimensional spatial distance between observation point i and observation point j; The digital elevation space distance between observation point i and observation point j; The time distance between observation point i and observation point j, λ, δ and μ are scale adjustment factors, which are used to balance the scale differences of different four-dimensional space-time distances. Here, u corresponds to the X axis of the coordinates in the figure, v corresponds to the Y axis of the coordinates in the figure, z corresponds to the Z axis of the coordinates in the figure, and t corresponds to the T axis of the coordinates in the figure.

In the kernel function, the selection of the bandwidth parameter has a great influence on the fitting result of the model. If the bandwidth is too small, the estimation result will be over-fitted. If the bandwidth is too large, the estimation result will be inaccurate. Therefore, the selection of appropriate bandwidth parameters is crucial to the accuracy of model estimation. Applied to this embodiment, that is to say, the four-dimensional space-time bandwidth h ^4D-ST is very important for the estimation accuracy of the 4D-GTWR model. The four-dimensional space-time bandwidth h ^4D-ST is obtained according to the two-dimensional space bandwidth h ^S2d , the digital elevation space bandwidth h ^SDEM and the time bandwidth h ^T .

Bring the d _ij ^ST described above into It can be concluded that:

Among them, w _ij ^S2d , w _ij ^SDEM , w _ij ^T represent the two-dimensional space weight, DEM space weight and time weight, respectively. Therefore, the four-dimensional space-time weight matrix of the 4D-GTWR model is expressed as follows:

From the above expression of W(u _i ,vi ,z _i ,t _{i )} , it can be seen that the four-dimensional space-time weight matrix is related to the two-dimensional space weight, DEM space weight and time weight. Need to solve h ^S2d , h ^SDEM , h ^T , λ, δ, μ.

In this embodiment, h ^S2d , h ^SDEM , h ^T , λ , δ and μ, as other embodiments, can also be calculated according to the CV criterion (cross-validation method), and the present invention does not limit the calculation process of the bandwidth and the scale adjustment factor.

The calculation process of AICc is:

σ ² =RSS _4D-GTWR /(n-2tr(S)+tr(S ^T S));

RSS _4D-GTWR = Y ^T (IS) ^T (IS) Y;

where ^σ2 is the unbiased estimate of the random error variance in the 4D-GTWR model, RSS _4D-GTWR is the residual sum of squares of the 4D-GTWR model, and I is the identity matrix.

The specific calculation process of RSS _4D-GTWR is:

in,

Hypothetical fit is an unbiased estimate of E(y _i ), that is but:

and E(ε ^T ε)=σ ² I. Then RSS _4D-GTWR can also be expressed as:

where the degree of freedom fd is n-2tr(S)+tr(S ^T S), so the unbiased variance estimate σ ² of the random error term in the 4D-GTWR model can be obtained as:

σ ² =RSS _4D-GTWR /(n-2tr(S)+tr(S ^T S)).

It can be seen from the above formula that the hat matrix S of the 4D-GTWR model is a matrix about the unknown parameters W(u _i ,vi ,z _i ,t _{i )} , then AICc is only about W(u _i ,vi ,z _{) i} , t _i ), but W(u _i ,vi ,z _i ,t _{i )} is a function of each bandwidth, so the final AICc is a function of each bandwidth, and the corresponding bandwidths and The scaling factor is each required bandwidth and scaling factor.

scaling factor λ, It is related to two-dimensional space, λ and δ are related to DEM space, and μ is related to time. In general, λ, Set it as a constant m. In the process of calculating δ, μ and each bandwidth, first set the range of δ, μ and each bandwidth (the basis for setting the bandwidth range is actually constantly tested in the calculation process, and the four-dimensional space-time distance).

The specific solution process is:

Step 1: Put λ and Set to m, then set the range of h ^S2d to determine the h ^S2d corresponding to the minimum AICc value;

Step 2: Put λ and Set to m, and then set the range of μ and h ^T to determine μ and h ^T corresponding to the minimum AICc value;

Step 3: Put λ and Set to m, and then set the range of δ and h ^SDEM to determine the δ and h ^SDEM corresponding to the minimum AICc value.

4) Through the above calculation, W(u _i ,vi ,z _i ,t _{i )} can be reversely solved, and then the hat matrix S of the 4D-GTWR model can be solved, and finally the prediction result can be obtained

Through the above calculations, the parameters in the 4D-GTWR model have been determined, and the 4D-GTWR model can finally be determined. When predicting PM2.5 concentration, the predicted meteorological data is brought into the 4D-GTWR model, and each observation can be predicted. PM2.5 concentration in the area.

The method is verified by the following specific historical data:

Take a certain observation point as an example to verify, let a certain observation point be (u ₁ , v ₁ , z ₁ , t ₁ ), and the other two observation points are (u ₂ , v ₂ , z ₂ , t ₂ ) and ( u ₃ , v ₃ , z ₃ , t ₃ ), the historical data of independent variables and dependent variables are shown in Table 1:

Table 1 Historical data

The above data is only a part of the historical data. In the process of determining the 4D-GTWR model, data of at least 100 observation points are required. Due to the excessive amount of data, they will not be listed here.

The space-time coordinates of the three observation points are shown in Table 2, and the two-dimensional position information is the abscissa and ordinate in the WGS_1984_World_Mercator projected coordinate system:

Table 2 Space-time coordinates of observation points

z t u v 161(z₁) 49(t₁) 12612250.93(u₁) 4111074.699(v₁) 147(z₂) 49(t₂) 12612591.07(u₂) 4114111.191(v₂) 131(z₃) 49(t₃) 12575408.2(u₃) 4105491.236(v₃)

Bring the historical data into the model, and do the correlation analysis to know the accuracy of the model. The correlation analysis includes R ² (correlation coefficient), RMSE (root mean square error), MAE (mean absolute error), R ² The larger the value, the smaller the RMSE and MAE, indicating that the model has high accuracy. Calculated as follows:

in, is the estimated value (predicted value) of PM2.5 concentration at observation point i, is the average value of PM2.5 concentration observations at observation point i, and y _i is the PM2.5 concentration observation value at observation point i.

Bring the above historical data into the GTWR model (the GTWR model is the prior art and will not be introduced too much here), and the results of the GTWR model will be obtained, including the independent variable regression coefficients of the independent variables in each observation point, and each observation point. The dependent variable regression fit vector of The bandwidth parameter is 2.2895, then the Perform linear correlation analysis with the actual vector value y (the PM2.5 observation value of each observation point), and the result is shown in Figure 2. The linear correlation analysis curve y=1.0536x-2.3442 (this curve is the correlation analysis curve) can be obtained. , x in this formula is the observed value of PM2.5, y is the predicted value of PM2.5), R ² , RMSE, MAE, the scattered points in Figure 2 have both the estimated value (predicted value) of PM2.5 and the The observed value of PM2.5 (measured value), the abscissa is PM2.5 (measured value), and the ordinate is the predicted value of PM2.5. The effect of the model can be seen through the relationship between the scattered points and the linear correlation analysis curve. The results are shown in Table 3:

Table 3 GTWR model

In Table 3, Min is the minimum value, LQ is the lower quartile, Med is the median value, UQ is the upper quartile, and Max is the maximum value, which describes the change of the regression coefficients of the respective variables with time and space, and the intercept is the regression coefficient constant term of the independent variable, AOD in the table represents the independent variable regression coefficient corresponding to the independent variable AOD, relative humidity represents the independent variable regression coefficient corresponding to the independent variable relative humidity, and air pressure represents the independent variable regression coefficient corresponding to the independent variable air pressure, Temperature represents the independent variable regression coefficient corresponding to the independent variable temperature, wind speed represents the independent variable regression coefficient corresponding to the independent variable wind speed, and wind direction represents the independent variable regression coefficient corresponding to the independent variable wind direction. The rainfall data is 0, so there is no correlation regression coefficient here.

Bring the above historical data into the 4D-GTWR model respectively, and the results of the 4D-GTWR model will be obtained, including the independent variable regression coefficient of the independent variable in each observation point, and the dependent variable regression fitting vector of each observation point. The bandwidth parameter is 0.0063, then the Perform linear correlation analysis with the actual vector value y (the PM2.5 observed value of each observation point), the result is shown in Figure 3, the linear correlation analysis curve y=1.0253x-1.2749, R ² , RMSE, MAE, As shown in Figure 3, the abscissa, ordinate and scattered points in Figure 3 and Figure 2 have the same meaning as the curve, so I won't introduce too much here. The results are shown in Table 4:

Table 4 4D-GTWR model

The meanings in Table 4 are the same as those in Table 3. It can be seen from the results that in the GTWR model, R ² =0.8811, RMSE=12.9579, MAE=9.5815, and in the 4D-GTWR model, R ² =0.9496, RMSE=8.5931, MAE = 5.9498, the R ² of the 4D ^- GTWR model is larger than that of the GTWR model, and the RMSE and MAE of the 4D-GTWR model are smaller than those of the GTWR model, indicating that the 4D-GTWR model is more accurate.

Example of PM2.5 concentration prediction device:

The PM2.5 concentration prediction device proposed in this embodiment includes a memory, a processor, and a computer program stored in the memory and running on the processor. The processor implements PM2.5 concentration prediction when executing the computer program. method.

The specific implementation process of the PM2.5 concentration prediction method has been introduced in the above-mentioned embodiment of the PM2.5 concentration prediction method, and will not be repeated here.

Claims (10)

1. A PM2.5 concentration prediction method is characterized by comprising the following steps:

acquiring PM2.5 concentration historical data, meteorological historical data and atmospheric aerosol optical thickness historical data;

acquiring two-dimensional position information, elevation information and sampling time information of an observation point;

determining parameters of a 4D-GTWR model according to the data, thereby determining the 4D-GTWR model, wherein the 4D-GTWR model is as follows:

wherein x is_ikP independent variables are the k independent variables of the observation point i, and the independent variables are meteorological data; y is_iThe dependent variable is the PM2.5 concentration data of the observation point i, and the parameter of the 4D-GTWR model is β_ik(u_i,v_i,z_i,t_i)、β_i0(u_i,v_i,z_i,t_i) And ξ_i，β_ik(u_i,v_i,z_i,t_i) The independent variable regression coefficient of the k independent variable of the observation point i is related to the space position of the observation point i β_i0(u_i,v_i,z_i,t_i) Constant term of independent variable regression coefficient for observation point i ξ_iIs the random error of observation point i; (u)_i,v_i,z_i,t_i) Is the four-dimensional space coordinate of observation point i, where (u)_i,v_i) Representing two-dimensional spatial coordinates, (z)_i) Representing the elevation space coordinate (t)_i) Representing a time coordinate; n is the number of observation points;

and predicting the concentration of PM2.5 according to the predicted meteorological data, the atmospheric aerosol optical thickness data and the determined 4D-GTWR model.

2. The PM2.5 concentration prediction method according to claim 1, wherein the method of calculating the regression coefficient of the independent variable includes: solving a least squares estimate of the independent variable regression coefficient by establishing an objective function of the observation point i, the objective function being:

wherein, w_ij ^STIs a kernel function between observation point i and observation point j;

least squares estimation of the independent variable regression coefficientsComprises the following steps:

wherein, W (u)_i,v_i,z_i,t_i) Is a four-dimensional space-time weight matrix composed of w_ij ^STAnd (3) forming a matrix, wherein X is an independent variable matrix, and Y is an observed value matrix of a dependent variable.

3. The PM2.5 concentration prediction method according to claim 2, wherein the kernel function is a Gaussian kernel function, and the formula is:

wherein h is^4D-STIs a four-dimensional space-time bandwidth, d_ij ^STIs the four-dimensional space-time distance between the observation point i and the observation point j.

4. The PM2.5 concentration prediction method according to claim 3, wherein the elevation information is digital elevation information, and the four-dimensional space-time distance between the observation point i and the observation point j is calculated by the following formula:

wherein,the two-dimensional space distance between the observation point i and the observation point j is obtained;the digital elevation space distance between the observation point i and the observation point j is obtained;the time distance between the observation point i and the observation point j is obtained; lambda, lambda,Delta and mu are scale adjustment factors used for balancing the scale difference of different four-dimensional space-time distances.

5. The PM2.5 concentration prediction method of claim 4, wherein the four-dimensional space-time bandwidth is a two-dimensional space bandwidth h based on^S2dDigital elevation space bandwidth h^SDEMAnd time bandwidth h^TAnd (4) obtaining the product.

6. The PM2.5 concentration prediction method according to claim 5, wherein the two-dimensional spatial bandwidth, the digital elevation spatial bandwidth and the time bandwidth are calculated according to the akage information amount criterion, and the akage information amount criterion AICc is:

wherein σ²Is an unbiased estimate of random error variance in the 4D-GTWR model, S is the cap matrix of the 4D-GTWR model, X₁、X₂、…、X_nAre row vectors consisting of the arguments of observation points 1, 2, …, n, respectively.

7. The PM2.5 concentration prediction method of claim 6, wherein the unbiased estimation of random error variance σ in the 4D-GTWR model²The calculation formula of (2) is as follows:

σ²＝RSS_4D-GTWR/(n-2tr(S)+tr(S^TS))；

wherein the RSS_4D-GTWRAs residues of the 4D-GTWR modelThe difference is the sum of squares.

8. The PM2.5 concentration prediction method according to claim 7, wherein the Residual Sum of Squares (RSS) of the residuals of the 4D-GTWR model_4D-GTWRThe calculation formula of (2) is as follows:

RSS_4D-GTWR＝Y^T(I-S)^T(I-S)Y；

wherein I is an identity matrix.

9. The PM2.5 concentration prediction method according to claim 1, wherein the meteorological data includes at least air pressure, air temperature, relative humidity, rainfall, wind speed, and wind direction.

10. A PM2.5 concentration prediction apparatus comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the PM2.5 concentration prediction method according to any one of claims 1 to 9 when executing the computer program.