CN105243444A - City monitoring station air quality prediction method based on online multi-core regression - Google Patents

Abstract

The present invention relates to a city monitoring station air quality prediction method based on online multi-core regression. The method comprises: firstly, based on historical data, extracting multi-field features of a city monitoring station, such as a forecasting meteorological feature, a real-time meteorological feature, a traffic feather, local and peripheral city air pollutant feature and the like; then, based on the extracted features, training a multi-core regression model, and performing online adjustment on the multi-core regression model by using new data; and finally, based on the adjusted model, predicting the air quality of the monitoring station hour by hour in a future period of time. The method can be used for accurately and efficiently predicting the air quality of the city monitoring station and has a guiding effect on environment protection and public life.

Description

A kind of city monitoring station Air Quality Forecast method returned based on online multinuclear

Technical field

The present invention relates to air quality monitoring field, particularly relate to a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear.

Background technology

Air is biological material of depending on for existence on the earth.Air quality and daily life closely bound up, in urban environmental qualitative assessment, occupy critical role.But along with human civilization and expanding economy, air pollution is more and more serious, how to improve air quality, Accurate Prediction air quality becomes more and more important.According to Air Quality Forecast result, people can take corresponding measure (as band mouth mask, avoiding going out) to avoid being subject to the infringement of air pollutants.On the other hand, environmental protection is the cause of the whole society, and the degree of participation of the public to environmental protection is that whether successfully important symbol is carried out in a national environmental protection job.If urban air-quality can announce forecast result every day as weather, the public just can understand the environmental quality truth in oneself living space, is conducive to the people and participates in and monitoring environment protection work.

The problem of conventional air qualitative forecasting method is mainly characteristic sum model two aspect:

From feature aspect, because air pollutants constantly can be flowed along with air, so the air pollutants level of surrounding cities is closely related, if the air quality in a ratio city is severe contamination and blows north wind, the air quality being so positioned at this Shelter in South China Cities, city after a period of time also can be affected.And traditional Air Quality Forecast method only considers the feature of the association areas such as meteorology, traffic, local air pollutant, do not consider that the Air Quality of surrounding cities treats the impact of predicted city, thus have impact on the accuracy predicted the outcome.On the other hand, traditional Air Quality Forecast method uses the correlated characteristic based on real time meteorological data in the model training stage, forecast period then uses the correlated characteristic based on forecast weather data, and it is all relevant to air quality with forecast weather data in real time, should give to consider simultaneously, otherwise the validity of model can be affected.

From model level, because production model (as Markov model etc.) has the inherent shortcoming such as marking bias and independence assumption, cause its predictablity rate not ideal; And discriminative model (as decision tree, support vector machine etc.) is although simpler than production model, but because its black box operates, the relation between demonstrating data can not be known, thus the characteristic of training data itself can not be reflected, and then negative effect is created to its predictive ability.Although conditional random field models had both had discriminative model than the advantage being easier to learn, the transition probability between contextual tagging can be considered again as production model, but it is the same with discriminative model with traditional production model, it is all the mode of learning of batch type, when there being new data, need to carry out re-training based on total data.Because re-training cost is high, model is made to be difficult to upgrade in time.Although online monokaryon returns the above-mentioned shortcoming that can overcome batch type transaction module, they are often before learning tasks, just secure a kernel function, if data stream is along with time instability change, will cause undesirable prediction effect.

Summary of the invention

The present invention overcomes above-mentioned weak point, object is to provide a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear, first the multi-field feature of air quality monitoring station's point is extracted based on historical data, as Meteorological Characteristics, traffic characteristic, this locality and surrounding cities air pollutants feature etc., then based on the features training multinuclear regression model extracted, and new data are utilized to carry out on-line tuning to multinuclear regression model; Finally based on the model after adjustment to the air quality in monitoring station following a period of time carry out by hour prediction.This method can predict air quality more accurately.

The present invention achieves the above object by the following technical programs: a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear, comprises the steps:

(1) pre-service is carried out to history raw data and obtain historical data sample, obtain training dataset (X based on historical data sample _k, Y _k) and core pond KP;

(2) combined training data set (X _k, Y _k) and core pond KP to submodule M ^ktrain, prediction of output model M={ M ^k| 1≤k≤h};

(3) utilize real time new Monitoring Data to each submodel M ^k(1≤k≤h) is adjusted to and forecast model M is updated to the forecast model after adjustment

(4) based on forecast model to future, the air quality of each moment p to be predicted is predicted.

As preferably, the history raw data of described step (1) comprises the meteorological correlated characteristic of forecast real-time weather correlated characteristic air pollutants correlated characteristic traffic correlated characteristic periphery monitoring station feature with surrounding cities feature

As preferably, described step (1) obtains training dataset (X _k, Y _k) and the method flow of core pond KP as follows:

1) alignment of data is carried out to history raw data, and complete data scrubbing with mean value replacement missing values and extremum, obtain historical data sample x={x _j| 1≤j≤n}, wherein, n represents number of samples, x _jit is the vector of the history raw data in jth hour and air quality composition in the past;

2) based on historical data sample, be submodel M ^kstructure training dataset:

2.1) feature composition training feature vector collection is extracted wherein $X_k^j=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

2.2) air quality of all samples forms flag sequence Y _k={ Y ^j| 1≤j≤n}, Y ^jrepresent sample x _jmark; And obtain X _kwith Y _kcomposition M ^ktraining dataset (X _k, Y _k);

3) h step 2 is repeated), obtain training dataset D={ (X _k, Y _k) | 1≤k≤h}, wherein h represents the time range that prediction is maximum;

4) choose the individual different kernel function of m and form core pond KP={kf _s| 1≤s≤m}, kf _srepresent s kernel function in core pond.

As preferably, the submodule M of described step (2) ^kdevice and weight vectors collection Weights is returned by m monokaryon ^kcomposition, wherein each monokaryon returns device by support vector collection parameter set form.

As preferably, the steps flow chart of described step (2) prediction of output model M is as follows:

I () trains s monokaryon to return device by simple regression error minimization algorithm, obtain support vector collection and parameter set

(ii) Repeated m time step (i), obtains the support vector collection that m monokaryon returns device and parameter set obtain submodel M ^ksupport vector collection parameter set ${\mathrm{Alpha}}^k=\left\{{\mathrm{Alpha}}_s^k\right|1\≤s\≤m\};$

(iii) weight vectors collection Weights is obtained by stochastic gradient descent algorithm ^k;

(iv) repeat h step (i) to step (iii), complete the training to h submodel; Obtain and prediction of output model M={ M ^k| 1≤k≤h}.

As preferably, described step (3) obtains with the forecast model after adjustment the method that Methods and steps (2) used is used is identical.

As preferably, the air quality of described step (4) to each moment p to be predicted in future is predicted, if the interval hourage of moment p to be predicted and current time is k, 1≤k≤h, method is as follows:

(I) based on forecast weather data and the historical data of moment p, feature composition vectorial X to be predicted is extracted ^* _k, $X_k^{*}=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

(II) by X ^* _kas forecast model input, obtain the predicted value Y of moment k to be predicted ^* _k;

(III) repeat h step (I) to step (II), complete the prediction to the following h air quality of individual hour;

(IV) prediction of output sequence Y ^*={ Y ^* _k| 1≤k≤h}.

Beneficial effect of the present invention is: (1) introduces the field correlated characteristics such as meteorology, traffic and local air pollutant, but also introduce surrounding cities air pollutants feature, consider that surrounding cities air quality treats the impact of predicted city air quality, thus air quality can be predicted more accurately; (2) consider real time meteorological data, also consider that forecast weather data is on the impact of air quality, makes forecast model more efficient simultaneously; (3) adopt online multinuclear regression model, not only overcome conventional batch formula transaction module and to upgrade in time the larger defect of cost, and solve online monokaryon and return performance issue because kernel function fixedly causes.

Accompanying drawing explanation

Fig. 1 is the city monitoring station Air Quality Forecast method flow diagram that the present invention is based on the recurrence of online multinuclear;

Fig. 2 is that partial process view set up by embodiment of the present invention model;

Fig. 3 is embodiment of the present invention kernel function details schematic diagram;

Fig. 4 is the adjustment of embodiment of the present invention on-time model and prediction process flow diagram.

Embodiment

Below in conjunction with specific embodiment, the present invention is described further, but protection scope of the present invention is not limited in this:

Embodiment: as shown in Figure 1, a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear, first obtains multinuclear regression model by model foundation part, secondly when there being new data, carrying out on-line tuning, obtain new model M to model; Then on-line prediction is carried out based on M.

The method is divided into two large divisions: model sets up part and online part.Wherein, model foundation part comprises data prediction and two stages of model training; Online part comprises data prediction, model adjustment and prediction three phases.Concrete implementation step is as follows:

Part set up by model:

Model foundation part mainly sets up forecast model M based on historical data sample.Because (the interval hourage of moment p and current time is k to following each moment p to be predicted, 1≤k≤h, h represents prediction maximum magnitude, when by hour prediction, k and h is integer, and unit is hour) air quality all by a corresponding submodel M ^kpredict, so M comprises h submodel.Its process flow diagram as shown in Figure 2.

Data preprocessing phase:

Step 1, for certain air quality monitoring stations point s in a city, alignment of data (namely with the unified each FIELD Data of a time and space unit) is carried out to the history raw data (as forecast weather data, real time meteorological data, traffic data, air pollutants data) in its each field, and data scrubbing (namely replacing missing values and extremum etc. with mean value), obtain historical data sample x={x _j| 1≤j≤n}, (x _jbe a vector be made up of forecast weather data, real time meteorological data, traffic data and the air pollutants data in website s place jth hour, n represents number of samples);

Step 2, based on historical data sample x, is submodel M ^kstructure training dataset:

1) feature composition training feature vector collection is extracted $X_k^j=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\},$ Wherein, represent forecast meteorological correlated characteristic, real-time weather correlated characteristic, air pollutants correlated characteristic, traffic correlated characteristic, periphery monitoring station characteristic sum surrounding cities feature respectively;

A) meteorological correlated characteristic is forecast

Forecast that meteorological correlated characteristic mainly considers that website s counts forecast weather data (as temperature, humidity, the wind-force) impact on its following air quality of the past period from the moment to be predicted, statistics feature (as maximal value, extreme difference, mean value, median and variance etc.) is extracted to forecast weather data;

B) real-time weather correlated characteristic

Real-time weather correlated characteristic mainly considers that website s counts real time meteorological data (as temperature, humidity, the wind-force) impact on its following air quality of the past period from the moment to be predicted, extracts statistics feature (as maximal value, extreme difference, mean value, median and variance etc.) to real time meteorological data;

C) air pollutants correlated characteristic

Air pollutants correlated characteristic mainly considers that website s counts air pollutants data (CO, NO of the past period from the current moment ₂, SO ₂, O ₃, PM _2.5, PM ₁₀) impact on its following air quality, statistics feature (as maximal value, extreme difference, mean value, median and variance etc.) is extracted to air pollutants data;

D) traffic correlated characteristic

Traffic correlated characteristic mainly considers traffic (as speed per hour, the traffic congestion index etc.) impact on its following air quality of the n bar section the past period near website s, extracts statistics feature (as expectation, variance etc.) to each section speed per hour;

E) periphery monitoring station feature

Periphery measuring station point patterns F _sconcrete extraction step be: count from current time, the air quality of each air quality monitoring stations point statistics feature in the past in l hour, as maximal value max _i, extreme difference r _i, mean value mean _i, median median _iand variance v _i, i.e. F _s={ max _i, r _i, mean _i, median _i, v _i| 1≤i<num}, wherein num is the quantity that this urban air-quality detects website.

F) surrounding cities feature

First, given city C to be predicted, for each prefecture-level city P of China _i, can P be obtained _iwith the angle of Distance geometry on wind rose of C.Based on this Distance geometry angle, the surrounding cities Q={q of the air quality affecting C of must be able to attending the meeting according to the wind direction of current time C _i| 1≤i≤count}, wherein count is the quantity of the surrounding cities affecting C air quality.

Surrounding cities feature F _cconcrete extraction step be: count from current time, the air quality in each city of the periphery statistics feature in the past in l hour, as maximal value max _i, extreme difference r _i, mean value mean _i, median median _iand variance v _i, i.e. F _s={ max _i, r _i, mean _i, median _i, v _i| 1≤i≤count}.

2) air quality of all samples forms flag sequence Y _k={ Y ^j| 1≤j≤n}, (Y ^jrepresent sample x _jair quality;

3) X _kwith Y _kcomposition M ^ktraining dataset (X _k, Y _k);

Step 3, repeats h step 2, obtains training dataset D={ (X _k, Y _k) | 1≤k≤h};

Step 4, by selecting different kernel function forms or arranging different parameters to a kernel function form, obtains core pond KP={kf _s| (m represents core pond Kernel Function number to 1≤s≤m}, kf _srepresent s kernel function in core pond), kernel function details are as shown in Figure 3.

The model training stage:

Step 1, based on the training dataset (X that pretreatment stage obtains _k, Y _k) and core pond KP, antithetical phrase model M ^k(primarily of SV ^k, Alpha ^kand Weights ^kcomposition) train:

1) train s monokaryon to return device by simple regression error minimization algorithm, obtain its support vector collection parameter set initialization with be sky; Then to data set X _kin each sample

A) in order to solve data aging problem, in reducing with multiplying power element, namely as shown in formula (1):

${\mathrm{Alpha}}_s^k=\mathrm{\&psi;}\&times;{\mathrm{Alpha}}_s^k---\left(1\right)$

Wherein, ψ is an adjustable parameter, 0 < ψ < 1;

B) based on the anticipation function of such as formula (2), s monokaryon can be obtained and return device to the predicted value of t sample

$F_s\left(X_k^t\right)={\mathrm{\&Sigma;}}_{i=1}^{t-1}{\mathrm{\&alpha;}}_i^s{\mathrm{kf}}_s(x_i,{\mathrm{SV}}_s^k)---\left(2\right)$

${\mathrm{\&alpha;}}_t^s=\mathrm{\&xi;}\&times;\mathrm{si}gn(Y^t-f_s(X_k^t\left)\right)---\left(3\right)$

Wherein, represent that s monokaryon returns the weight of i-th support vector of device, calculated by formula (3), formula (1) upgrades; ξ represents an adjustable parameter; When time, ${\mathrm{SV}}_s^k=\{{\mathrm{SV}}_s^k\&cup;t\},{\mathrm{Alpha}}_s^k=\{{\mathrm{Alpha}}_s^k\&cup;{\mathrm{\&alpha;}}_t^s\}.$

2) step 1 is repeated), obtain m monokaryon recurrence device with and then can obtain ${\mathrm{SV}}^k=\left\{{\mathrm{SV}}_s^k\right|1\&le;s\&le;m\},{\mathrm{Alpha}}^k=\{{\mathrm{Alpha}}_s^k|1\&le;s\&le;m\};$

3) obtain by stochastic gradient descent algorithm the weight vectors collection Weights combining multinuclear ^k; Initialization Weights ^kfor m ties up null vector; To each data sample adopt the update rule as formula (4):

${\mathrm{Weights}}^k\&LeftArrow;{\mathrm{Weights}}^k-\mathrm{\&eta;}\&times;(Y^t-{\left({\mathrm{Weights}}^k\right)}^{T}f(X_k^t\left)\right)f\left(X_k^t\right)---\left(4\right)$

Wherein, η is learning rate parameter, f (x)={ f _s(x) | 1≤s≤m}, $f_s\left(X_k^t\right)={\left({\mathrm{Alpha}}^k\right)}^T{\mathrm{kf}}_s(X_k^t,{\mathrm{SV}}^k);$

Step 2, repeats h step 1, completes the training to h submodel;

Step 3, prediction of output model M={ M ^k| 1≤k≤h}.

Online part:

Online part is after obtaining multinuclear regression model, whenever having new data, according to these data by each submodel M ^k(1≤k≤h) on-line tuning is and forecast model M is before updated to the forecast model after adjustment forecast period is based on each submodel M ^kto future each moment p to be predicted (the interval hourage of moment p and current time is that k, 1≤k≤h, h represents prediction maximum magnitude, when by hour prediction, k and h is integer, and unit is hour) air quality predict, as shown in Figure 4.

Data preprocessing phase:

Step 1, each FIELD Data (as forecast weather data, real time meteorological data, traffic, air pollutants) that current time i newly inputs is alignd (namely with the unified each FIELD Data of a time and space unit), and data scrubbing (namely by the missing values in mean value replacement raw data and extremum), obtain data sample

Step 2, according to cleaned data sample and historical data sample, be submodel M ^kstructure adjustment data set:

1) 6 category features described in extraction form a training feature vector collection ${\hat{X}}_k=\left\{\widehat{X_k^i}\right|1\&le;k\&le;h\},\widehat{X_k^i}=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

2) data sample corresponding air quality form flag sequence

3) with composition M ^kadjustment data set

Step 3, repeats h step 2, is adjusted data set the model adjusting stage:

Step 1, based on the adjustment data set that pretreatment stage obtains the core pond KP in part pretreatment stage step 4 is set up, by M with model ^kon-line tuning is (primarily of and composition):

1) adjust s monokaryon by simple regression error minimization algorithm and return device, namely upgrade its support vector collection be updated to by parameter set be updated to

First in order to solve data aging problem, in reducing with multiplying power element, namely ${\mathrm{Alpha}}_s^k=\mathrm{\&psi;}\&times;{\mathrm{Alpha}}_s^k(0<\mathrm{\&psi;}<1);$ Then can obtain according to formula (3) if time, $\widehat{{\mathrm{SV}}_s^k}=\{{\mathrm{SV}}_s^k\&cup;(n+1)\},\widehat{{\mathrm{Alpha}}_s^k}=\{{\mathrm{Alpha}}_s^k\&cup;{\mathrm{\&alpha;}}_{n+1}^s\};$

2) Repeated m time step 1), obtains m monokaryon recurrence device with and then can obtain $\widehat{{\mathrm{SV}}^k}=\left\{\widehat{{\mathrm{SV}}_s^k}\right|1\&le;s\&le;m\},\widehat{{\mathrm{Alpha}}^k}=\{\widehat{{\mathrm{Alpha}}_s^k}|1\&le;s\&le;m\};$

3) by the weight vectors collection Weights of stochastic gradient descent algorithm more Combination nova multinuclear ^kfor

Can be obtained by formula (4):

$\widehat{{\mathrm{Weighs}}^k}\&LeftArrow;{\mathrm{Weights}}^k-\mathrm{\&eta;}\&times;(\widehat{Y_k}-{\left({\mathrm{Weights}}^k\right)}^{T}f(\widehat{X_k}\left)\right)f\left(\widehat{X_k}\right)---\left(5\right)$

Wherein, η is learning rate parameter, f (x)={ f _s(x) | 1≤s≤m}, $f_s\left(\widehat{X_k}\right)={\left(\widehat{{\mathrm{Alpha}}_k}\right)}^T{\mathrm{kf}}_s(\widehat{X_k},\widehat{{\mathrm{SV}}_k});$

Step 2, repeats h step 1, completes the adjustment to h submodel, obtains forecast model $\hat{M}=\left\{\widehat{M^k}\right|1\&le;k\&le;h\};$

Step 3, is updated to the forecast model after adjustment by forecast model M before and data sample number n is updated to n+1.

Forecast period:

Step 1, based on forecast model the air quality in moment p to be predicted in future (wherein, the interval hourage of moment p and current time is k, 1≤k≤h) is predicted:

A) based on forecast weather data and the historical data of moment p, 6 category feature composition vectorial X to be predicted is extracted ^* _k, $X_k^{*}=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

B) by X ^* _kforecast model is obtained as the model adjusting stage input, the predicted value Y of moment k to be predicted can be obtained ^* _k, namely in formula (8)

$F\left(X_k^{*}\right)={\left({\mathrm{Weights}}^k\right)}^{T}f\left(X_k^{*}\right)---\left(8\right)$

Wherein, f (x)={ f _s(x) | 1≤s≤m}, $f_s\left(X_k^{*}\right)={\left({\mathrm{Alpha}}^k\right)}^T{\mathrm{kf}}_s(X_k^{*},{\mathrm{SV}}^k);$

Step 2, repeats h step 1, completes the prediction to the following h air quality of individual hour;

Step 3, prediction of output sequence Y ^*={ Y ^* _k| 1≤k≤h}.

The know-why being specific embodiments of the invention and using described in above, if the change done according to conception of the present invention, its function produced do not exceed that instructions and accompanying drawing contain yet spiritual time, must protection scope of the present invention be belonged to.

Claims (7)

1., based on the city monitoring station Air Quality Forecast method that online multinuclear returns, it is characterized in that, comprise the steps:

(1) pre-service is carried out to history raw data and obtain historical data sample, obtain training dataset (X based on historical data sample _k, Y _k) and core pond KP;

(2) combined training data set (X _k, Y _k) and core pond KP to submodule M ^ktrain, prediction of output model M={ M ^k| 1≤k≤h};

(3) utilize real time new Monitoring Data to each submodel M ^k(1≤k≤h) is adjusted to and forecast model M is updated to the forecast model after adjustment

(4) based on forecast model to future, the air quality of each moment p to be predicted is predicted.

2. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1, is characterized in that, the history raw data of described step (1) comprises the meteorological correlated characteristic of forecast real-time weather correlated characteristic air pollutants correlated characteristic traffic correlated characteristic periphery monitoring station feature with surrounding cities feature

3. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1 and 2, it is characterized in that, described step (1) obtains training dataset (X _k, Y _k) and the method flow of core pond KP as follows:

1) alignment of data is carried out to history raw data, and complete data scrubbing with mean value replacement missing values and extremum, obtain historical data sample x={x _j| 1≤j≤n}, wherein, n represents number of samples, x _jit is the vector of the history raw data in jth hour and air quality composition in the past;

2) based on historical data sample, be submodel M ^kstructure training dataset:

2.1) feature composition training feature vector collection is extracted wherein $X_k^j=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

2.2) air quality of all samples forms flag sequence Y _k={ Y ^j| 1≤j≤n}, Y ^jrepresent sample x _jmark; And obtain X _kwith Y _kcomposition M ^ktraining dataset (X _k, Y _k);

3) h step 2 is repeated), obtain training dataset D={ (X _k, Y _k) | 1≤k≤h}, wherein h represents the time range that prediction is maximum;

4) choose the individual different kernel function of m and form core pond KP={kf _s| 1≤s≤m}, kf _srepresent s kernel function in core pond.

4. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1, is characterized in that, the submodule M of described step (2) ^kdevice and weight vectors collection Weights is returned by m monokaryon ^kcomposition, wherein each monokaryon returns device by support vector collection parameter set form.

5. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1 or 4, it is characterized in that, the steps flow chart of described step (2) prediction of output model M is as follows:

I () trains s monokaryon to return device by simple regression error minimization algorithm, obtain support vector collection and parameter set

(ii) Repeated m time step (i), obtains the support vector collection that m monokaryon returns device and parameter set obtain submodel M ^ksupport vector collection parameter set ${\mathrm{Alpha}}^k=\left\{{\mathrm{Alpha}}_s^k|1\&le;s\&le;m\right\};$

(iii) weight vectors collection Weights is obtained by stochastic gradient descent algorithm ^k;

(iv) repeat h step (i) to step (iii), complete the training to h submodel; Obtain and prediction of output model M={ M ^k| 1≤k≤h}.

6. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1, it is characterized in that, described step (3) obtains with the forecast model after adjustment the method that Methods and steps (2) used is used is identical.

7. a kind of city monitoring station Air Quality Forecast method returned based on online multinuclear according to claim 1, it is characterized in that: the air quality of described step (4) to each moment p to be predicted in future is predicted, if the interval hourage of moment p to be predicted and current time is k, 1≤k≤h, method is as follows:

(I) based on forecast weather data and the historical data of moment p, feature composition vectorial X to be predicted is extracted ^* _k, $X_k^{*}=\{F_k^p,F_k^m,F_k^a,F_k^t,F_k^s,F_k^c\};$

(II) by X ^* _kas forecast model input, obtain the predicted value Y of moment k to be predicted ^* _k;

(III) repeat h step (I) to step (II), complete the prediction to the following h air quality of individual hour;

(IV) prediction of output sequence Y ^*={ Y ^* _k| 1≤k≤h}.