KR20210086786A - System and method for predicting fine dust and odor - Google Patents

Abstract

The present invention relates to a system for predicting fine dust and odors. The system comprises: a time series prediction model for acquiring and outputting estimated air quality values corresponding to air quality information and weather information by a deep learning method; a spatial attribute-based prediction model for acquiring and outputting, in regional units, the spatial variation values of the air quality information through a geographically weighted regression (GWR) model; a space-time prediction model for acquiring and outputting the finally estimated air quality values corresponding to outputs of the time series prediction model and the spatial attribute-based prediction model by the deep learning method; and an air quality prediction unit which first acquires the estimated air quality values of an area to be estimated and the spatial variation values of the air quality information through the time series prediction model and the spatial attribute-based prediction model if the area to be predicted is selected, and then finally acquires and outputs the finally estimated air quality values corresponding to the estimated air quality values and the spatial variation values of the air quality information through the time series prediction model. The air quality information includes fine dust information and odor information. Therefore, the system can perform a prediction operation considering the time series attributes and spatial attributes of air quality.

Description

System and method for predicting fine dust and odor

The present invention relates to a system and method for predicting fine dust and odor in a new way that can more accurately predict air quality in an area where there is no air quality measuring station, and can analyze and predict both fine dust and odor.

Most of the air quality prediction methods currently used in the field of weather forecasting are based on atmospheric diffusion models to consider pollution sources, topography, wind direction, and wind speed.

This type of model makes predictions in consideration of atmospheric changes from the point of occurrence of pollutants, so there is a fundamental limit to reflect changes in predicted values such as fine dust and odors according to spatial characteristics.

On the other hand, recently, time series models using machine learning have been developed, which also have similar limitations in that spatial information is neglected.

In particular, in the case of an urban area, there is a very high possibility that a clear difference in air quality will occur even in adjacent areas. For example, as shown in FIG. 1 and Table 1, the difference in PM10 of fine dust at two measuring stations (cells 12 and 1) that are about 1 km apart is almost double, and the complex odor (OU) of cells 12 and 22, which are immediately adjacent, : It can be seen that the difference between the weighted sum of _{H 2 S, NH3 and VOS) can be about 19 times.}

Cell No.
PM2.5
PM10
OU
2
52.53
311.67
0.02
12
50.74
*184.94
*0.11
22
49.82
127.66
*1.95
13
87.84
*92.27
0.06
41
57.21
147.15
0.03
46
62.49
90.05
2.20
74
42.80
117.83
0.00
80
51.05
134.66
0.01

As such, it can be seen that the air quality concentration in an urban area is particularly greatly affected by spatial characteristics. Therefore, it can be inferred that it is necessary to basically reflect spatial information in the air quality prediction in an area where there is no air quality measurement station.

Accordingly, in order to solve the above problems, the present invention is to provide a system and method for predicting fine dust and odors that can perform a prediction operation considering both the time series and spatial properties of air quality.

In addition, it is intended to provide a fine dust and odor prediction system and method that can analyze and predict both fine dust and odor.

The object of the present invention is not limited to the object mentioned above, and other objects not mentioned will be clearly understood by those of ordinary skill in the art to which the present invention belongs from the following description.

As a means for solving the above problem, according to an embodiment of the present invention, a time series prediction model for obtaining and outputting air quality information and air quality prediction values corresponding to weather information in a deep learning method; a spatial attribute-based prediction model for obtaining and outputting spatial variation values of the air quality information in units of regions through a geographic weighted regression model (GWR); a spatiotemporal prediction model for acquiring and outputting air quality final prediction values corresponding to outputs of the time series prediction model and the spatial property-based prediction model in a deep learning method; and when a prediction target area is selected, the air quality prediction value of the prediction target area and the spatial variation value of air quality information are first obtained through the time series prediction model and the spatial attribute-based prediction model, and then through the spatio-temporal prediction model An air quality prediction unit that finally acquires and outputs a final air quality prediction sheet corresponding to the air quality prediction value and the spatial variation value of the air quality information, wherein the air quality information is composed of fine dust information and odor information It provides a dust and odor concentration prediction system.

The fine dust and odor concentration prediction system learns the correlation between the air quality information and weather information and the air quality prediction model to the time series prediction model, and then outputs the time series prediction model and the spatial attribute-based prediction model in succession. And it characterized in that it further comprises a prediction model learning unit for additionally learning the correlation between the air quality final prediction value to the spatiotemporal prediction model.

The time series prediction model is characterized in that it is implemented as a Long Short Term Memory (LSTM) of a recurrent neural network (RNN).

The fine dust and odor concentration prediction system acquires and stores air quality information, weather information, and spatial information in units of regions, and after performing regional clustering based on spatial properties, air quality information measured through an air quality measuring station and It characterized in that it further comprises a data collection unit for sharing the weather information in a cluster unit.

As a means for solving the above problem, according to another embodiment of the present invention, a time series prediction model for acquiring and outputting air quality information and air quality prediction values corresponding to weather information in a deep learning method; a spatial attribute-based prediction model for obtaining and outputting spatial variation values of the air quality information in units of regions through a geographic weighted regression model (GWR); And the method of predicting fine dust and odor concentration using an ensemble prediction model having a spatio-temporal prediction model for obtaining and outputting final air quality prediction values corresponding to the outputs of the time series prediction model and the spatial property-based prediction model in a deep learning method is Iteratively learns the correlation between air quality information and weather information and the air quality prediction model in the time series prediction model, and at the same time, the correlation between the output of the time series prediction model and the spatial attribute-based prediction model and the final air quality prediction value repeatedly learning the spatiotemporal prediction model; and when a prediction target area is selected, the air quality prediction value of the prediction target area and the spatial variation value of air quality information are first obtained through the time series prediction model and the spatial attribute-based prediction model, and then through the spatio-temporal prediction model It may include the step of finally obtaining and outputting the final air quality predicted value corresponding to the air quality predicted value and the spatial variation value of the air quality information.

The present invention newly proposes an ensemble model combining a time series model and a spatial property-based model for accurate concentration prediction in a stacking structure, and through this, the time series property of air quality is corrected according to the spatial property, and the final air quality prediction value is obtained and print it out. In other words, it is possible to predict air quality change patterns that sufficiently consider spatial properties as well as time series properties, and as a result, it is possible to more accurately predict air quality in areas without air quality monitoring stations.

In addition, in the present invention, it is also possible to obtain and provide air quality prediction information having more diverse information by analyzing and predicting fine dust and odors together.

1 is a diagram for explaining air quality measurement results that are different for each region of an urban area.
2 is a view for explaining a system for predicting concentrations of fine dust and odors according to an embodiment of the present invention.
3 is a diagram for explaining the concept of a method for predicting concentrations of fine dust and odors according to an embodiment of the present invention.
4 is a diagram for explaining a time series prediction model according to an embodiment of the present invention.
5 is a diagram for explaining a regional clustering method for acquiring regional air quality and weather information according to an embodiment of the present invention.

The following is merely illustrative of the principles of the invention. Therefore, those skilled in the art will be able to devise various devices which, although not explicitly described or shown herein, embody the principles of the present invention and are included within the spirit and scope of the present invention. Moreover, it is to be understood that all conditional terms and examples listed herein are, in principle, expressly intended solely for the purpose of enabling the concept of the present invention to be understood, and not limited to the specifically enumerated embodiments and states as such. should be

Moreover, it is to be understood that all detailed description reciting the principles, aspects, and embodiments of the invention, as well as specific embodiments, are intended to cover structural and functional equivalents of such matters. It is also to be understood that such equivalents include not only currently known equivalents, but also equivalents developed in the future, i.e., all devices invented to perform the same function, regardless of structure.

Thus, for example, the block diagrams herein are to be understood as representing conceptual views of illustrative circuitry embodying the principles of the present invention. Similarly, all flowcharts, state transition diagrams, pseudo code, etc. may be tangibly embodied on computer-readable media and be understood to represent various processes performed by a computer or processor, whether or not a computer or processor is explicitly shown. should be

The functions of the various elements shown in the figures including a processor or functional blocks represented by similar concepts may be provided by the use of dedicated hardware as well as hardware having the ability to execute software in association with appropriate software. When provided by a processor, the functionality may be provided by a single dedicated processor, a single shared processor, or a plurality of separate processors, some of which may be shared.

In addition, clear use of terms presented as processor, control, or similar concepts should not be construed as exclusively referring to hardware having the ability to execute software, and without limitation, digital signal processor (DSP) hardware, ROM for storing software. It should be understood to implicitly include (ROM), RAM (RAM) and non-volatile memory. Other common hardware may also be included.

In the claims of the present specification, a component expressed as a means for performing the function described in the detailed description includes, for example, a combination of circuit elements that perform the function or software in any form including firmware/microcode, etc. It is intended to include all methods of performing the functions of the device, coupled with suitable circuitry for executing the software to perform the functions. Since the present invention defined by these claims is combined with the functions provided by the various enumerated means and combined in a manner required by the claims, any means capable of providing the functions are equivalent to those contemplated from the present specification. should be understood as

The above-described objects, features and advantages will become more apparent through the following detailed description in relation to the accompanying drawings, and accordingly, those of ordinary skill in the art to which the present invention pertains can easily implement the technical idea of the present invention. There will be. In addition, in describing the present invention, if it is determined that a detailed description of a known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted.

2 is a view for explaining a system for predicting concentrations of fine dust and odors according to an embodiment of the present invention.

As shown in FIG. 2, the system of the present invention generates two prediction values through a time series prediction model 111 for predicting time series properties of air quality and a spatial property based prediction model 112 for predicting spatial properties of air quality. After that, the ensemble prediction model 110 of the stacking type using these as an input of the spatio-temporal prediction model 113 again, a data collection unit that additionally acquires and stores spatial information in addition to air quality information and weather information of each region (120), the time-series prediction model 111 and the spatial property-based prediction model 112 are simultaneously trained through the air quality information, weather information, and spatial information of each region, and the spatial-temporal prediction model 113 using their output values ), when the prediction model learning unit 130 repeatedly performs the process of additional training, and the prediction target area is selected, the time series prediction model 111 and the space through air quality information, weather information, and spatial information of the prediction target area It is composed of an attribute-based prediction model 112 and an air quality prediction unit 140 that obtains and outputs the final air quality prediction site using sequentially through the spatio-temporal prediction model 113 .

That is, the present invention generates two prediction values through a time-series prediction model 111 for predicting time-series properties of air quality and a spatial-property-based prediction model 112 for predicting spatial properties of air quality, and then adds them to the spatio-temporal prediction model ( 113), the ensemble prediction model 110 of the stacking type is implemented to perform an air quality prediction operation considering not only the time series property of air quality but also the spatial property.

In other words, the present invention mixes the air quality time series prediction method and the spatial property-based prediction method using a geographically weighted regression model as shown in FIG. 3 to accurately predict the air quality in an area without an air measuring station.

Hereinafter, the function and operation of each component will be described in more detail as follows.

The time series prediction model 111 is for acquiring and outputting air quality prediction values corresponding to air quality and weather information (that is, a pattern in which air quality information changes with time according to weather information), and as shown in FIG. 4 , input It can be implemented as a Long Short Term Memory (LSTM) of a Recurrent Neural Network (RNN) composed of a layer, a hidden layer, and an output layer.

Air quality information is defined by a total of five properties: two types of fine dust (PM2.5, PM10), hydrogen sulfide (H ₂ S), which is a odor-causing substance, ammonia (NH ₃ ), and volatile organic compounds (VOC). It is preferable that the information is defined as a total of two properties, namely wind speed and wind direction, but of course, this may be variously changed in actual application examples in the future.

Accordingly, as shown in Table 2, the input layer has 7 nodes for receiving air quality information and weather information, the output layer has 5 nodes for outputting air quality forecasts excluding the weather information, and the hidden layer has 8 or 10 nodes. It can have multiple nodes.

In addition, hyperbolic tangent (tanh), logistic sigmoid (logistic sigmoid), and softmax (softmax) are applied as the activation functions of the input layer and hidden node state, the three gates, and the output layer, respectively, and the epoch ) and mini batch size can be set to 10 and 100. However, it is of course natural that the limitations may also be variously changed in actual application examples in the future.

Parameters Description No. of Inputs 7 (PM2.5, PM10, H ₂ S,NH ₃ ,WindDirection,WinSpeed) No. of hidden nodes 10 No. of output nodes 5 Input activation function tanh Gate activation function logistic sigmoid Hidden state activation function tanh Output activation function softmax Mini batch size 100 No. of epochs 10

The spatial attribute-based prediction model 112 is for acquiring and outputting spatial variation values of air quality information in units of regions, which can be implemented as a geographically weighted regression model (GWR), which is one of the regression models that analyze spatial variation. .

The key to geographic weighted regression is to treat the regression coefficients as a function of location, not a constant. In other words, the regression coefficient is not the same in all regions, but is modeled so that it varies according to spatial location. When there are y dependent variables and j independent variables x for a known position (u _i , v _i ), and the regression coefficients and error terms are β and ε, respectively, the regression equation can be expressed as Equation 1.

Here, y _i is the dependent variable of the i-th region, x _ij is the j-th explanatory variable in the i-th region, (u _i , v _i ) is the spatial coordinate of the i-th region, and β ₀ (u _i , v _i ) is The intercept coefficient, β _j (u _i , v _i ) is the j-th regression coefficient of the i-th region.

의 추정은 다음의 수학식2로 계산되어진다.Unlike the least-squares estimation model, the characteristic of this model is that it can identify regional influences by estimating individual regression coefficients locally. In order to estimate the regression coefficient of the model, it is assumed that the closer the distance to the geographic location i is, the more spatially it is affected than the farther away. Therefore, the regression coefficient

Estimation of is calculated by Equation 2 below.

At this time, each element of the geographic weight matrix W is calculated according to a kernel that is a weight function.

The kernel may take various forms, but Equation 3, which is generally a Gaussian form, is used.

At this time, in spatial coordinates (u _i , v _i ), the weight, d _i , (u _i , v _{i ) for the measured value i} is the distance between the measured value i and the spatial coordinates (u _i , v i ), and h is the bandwidth.

The kernel has a fixed method (Fixed spatial kernel) in which the bandwidth for making weights is fixed and a variable method (Adaptive spatial kernel) in which a different bandwidth is applied according to the number of measurements. When the measured values are regularly distributed in the survey area, the fixed method is used, and when the distribution of the measured values around the survey area is varied, the variable method is used. However, when the judgment is not certain, the variable method is usually adopted. In addition, Akaike Information Criterion (AIC) or Cross Validation (CV) is used to set the appropriate bandwidth, and the AIC method considering the difference between the measured value and the estimate and the complexity of the model is preferred.

It is preferable that the input variables of the geographically weighted regression model are spatial properties and air quality measurements as shown in Table 3, and each model was created for m pieces of air quality information.

The spatial attributes that can be collected by region include the number of factories, the number of cases under construction, the number of buildings, the average building area, green area, park area, residential area, commercial district area, road length, road width, number of intersections, number of residents, etc. However, in the present invention, by applying a correlation-based feature selection method using spatial attribute data and air quality measurements, five variables were finally selected and used.

That is, the geographic weighted regression model (GWR) of the present invention receives spatial information and air quality information of each region, and obtains and outputs spatial variation values of air quality information in consideration of spatial properties of each region.

Attributes Mean Min Max Standard Deviation Green Zone Size 109310 0 250000 98469 Residential Area Size 16107 0 132301 31733 Road Length 1185 0 4908 1117 Average Road Width 14229 0 49259 12825 Residential Population 430 0 3977 906

The spatio-temporal prediction model 113 is for acquiring and outputting air quality final prediction values corresponding to the outputs of the time-series prediction model and the spatial property-based prediction model in a deep learning method, and the time-series prediction model 111 and the spatial property-based prediction model After generating data in units of time, the result of (112) is used again as an input value for the LSTM model to calculate the final air quality forecast.

The spatiotemporal prediction model 113 is also implemented as an LSTM of a recurrent neural network, but the number of input nodes is 2xm, consisting of m output values of the time series model and m output values of the spatial property-based prediction model, and the number of hidden nodes is 2xl. configured, and the remaining parameters may be set the same as in Table 1 described above. The model 113 can generate and provide forecasts in hours and days, but is not limited thereto.

The data collection unit 120 acquires, stores, and manages air quality information, weather information, and spatial information in units of regions. That is, after the entire analysis target region is divided into a plurality of regions, air quality and weather information and spatial attribute information are respectively acquired and stored in units of regions.

However, it is practically impossible to directly acquire air quality and meteorological information for each of a plurality of regions by installing air quality measuring stations in all of a plurality of regions.

Accordingly, in the present invention, as in FIG. 5 , the spatial properties of the air quality measuring station and the spatial properties of the remaining cell areas are compared and analyzed, the cell areas having similar spatial characteristics are clustered into the same group, and then the measured values of the air quality measuring station are clustered. Allow sharing in units (red, green, yellow, blue). That is, a cell area without an air quality measuring station can receive and use the measured value of an air quality measuring station installed in a cell area having spatial characteristics similar to itself.

In this case, clustering is performed in a K-means method, and the number of clusters may be determined and adjusted through an Elbow algorithm, but is not limited thereto.

The predictive model learning unit 130 first trains the time series predictive model 111 and the spatial attribute-based predictive model 112, and then uses their output values to further train the space-time predictive model 113. By iteratively performing, the ensemble prediction model 110 can perform an air quality prediction operation that additionally considers spatial information in addition to air quality information and weather information.

In more detail, first, the predictive model learning unit 130 receives primary learning data having air quality information and weather information as input conditions and air quality prediction values as output conditions based on the information stored in the data collection unit 120 . Create multiple.

And by repeatedly learning the time series prediction model 111 through the primary learning data and simultaneously providing the air quality information provided to the time series prediction model 111 to the spatial property-based prediction model 112, the time series prediction model 111 ) to simultaneously acquire the spatial variation value of air quality information corresponding to the output value of .

In addition, secondary learning data having the air quality prediction value of the time series prediction model 111 and the spatial variation value of the air quality information of the spatial attribute-based prediction model 112 as input conditions, and the final air quality prediction value as output conditions are additionally generated and iteratively machine-learning the spatiotemporal prediction model 113 through these.

And when the output deviation of the spatiotemporal prediction model 113 converges to less than a preset value, the prediction model learning is terminated.

When any one of a plurality of regions is determined as a prediction target region, the prediction unit 140 acquires current air quality information, weather information, and spatial attribute information of the prediction target region through the data collection unit 120 .

At this time, if an air quality measuring station exists in the prediction target area, the air quality information and weather information of the corresponding air quality measuring station are used, but if there is no air quality measuring station, the air quality present in the cluster to which the prediction target area belongs Make it possible to share and use the measurement values of the measuring stations.

Then, air quality information and weather information are input to the time series prediction model 111, and air quality information and spatial property information are input to the spatial property based prediction model 112, and the time series prediction model 111 and the spatial property based prediction model Through (112), the air quality prediction value and the spatial variation value of air quality information are simultaneously acquired.

And by inputting the spatial variation value of the air quality prediction value and the air quality information back into the spatio-temporal prediction model 113, the spatio-temporal prediction model 113 obtains the final air quality prediction value considering all of the air quality information, weather information and spatial information, and to output.

The method according to the present invention described above may be produced as a program to be executed by a computer and stored in a computer-readable recording medium. Examples of the computer-readable recording medium include ROM, RAM, CD-ROM, magnetic tape. , a floppy disk, an optical data storage device, and the like, and also includes those implemented in the form of a carrier wave (eg, transmission through the Internet).

The computer-readable recording medium is distributed in network-connected computer systems, so that the computer-readable code can be stored and executed in a distributed manner. In addition, functional programs, codes, and code segments for implementing the method can be easily inferred by programmers in the art to which the present invention pertains.

In the above, preferred embodiments of the present invention have been illustrated and described, but the present invention is not limited to the specific embodiments described above, and it is common in the technical field to which the present invention pertains without departing from the gist of the present invention as claimed in the claims. Various modifications may be made by those having the knowledge of, of course, and these modifications should not be individually understood from the technical spirit or perspective of the present invention.

Claims (5)

a time series prediction model for acquiring and outputting air quality prediction values corresponding to air quality information and weather information in a deep learning method;
a spatial attribute-based prediction model for obtaining and outputting spatial variation values of the air quality information in units of regions through a geographic weighted regression model (GWR); and
a spatiotemporal prediction model for acquiring and outputting air quality final prediction values corresponding to the outputs of the time series prediction model and the spatial property-based prediction model in a deep learning method;
When the prediction target area is selected, the air quality prediction value of the prediction target area and the spatial variation value of the air quality information are first obtained through the time series prediction model and the spatial property-based prediction model, and then the air quality through the spatio-temporal prediction model It includes an air quality prediction unit that finally acquires and outputs the final prediction of air quality corresponding to the spatial variation of the quality prediction value and air quality information,
The air quality information
Fine dust and odor concentration prediction system, characterized in that it consists of fine dust information and odor information. According to claim 1,
After learning the correlation between the air quality information and the weather information and the air quality prediction model in the time series prediction model, the correlation between the output of the time series prediction model and the spatial attribute-based prediction model and the final air quality prediction value continuously thereafter Fine dust and odor concentration prediction system, characterized in that it further comprises a prediction model learning unit for additionally learning the spatio-temporal prediction model. According to claim 2, wherein the time series prediction model
A system for predicting the concentration of fine dust and odors, characterized in that it is implemented with LSTM (Long Short Term Memory) of a recurrent neural network (RNN). 3. The method of claim 2,
Data that acquires and stores air quality information, meteorological information, and spatial information in units of regions, but performs regional clustering based on spatial attributes, and then shares air quality information and weather information measured through air quality measurement stations in cluster units Fine dust and odor concentration prediction system, characterized in that it further comprises a collection unit. a time series prediction model for acquiring and outputting air quality prediction values corresponding to air quality information and weather information in a deep learning method; a spatial attribute-based prediction model for obtaining and outputting spatial variation values of the air quality information in units of regions through a geographic weighted regression model (GWR); And fine dust and odor concentration prediction method using an ensemble prediction model having a space-time prediction model that acquires and outputs the final air quality prediction value corresponding to the outputs of the time-series prediction model and the spatial property-based prediction model in a deep learning method ,
Iteratively learns the correlation between the air quality information and weather information and the air quality prediction model in the time series prediction model, and at the same time, the correlation between the output of the time series prediction model and the spatial attribute-based prediction model and the air quality final prediction value repeatedly learning the spatiotemporal prediction model; and
When the prediction target area is selected, the air quality prediction value of the prediction target area and the spatial variation value of the air quality information are first obtained through the time series prediction model and the spatial property-based prediction model, and then the air quality through the spatio-temporal prediction model A method for predicting the concentration of fine dust and odors, comprising the step of finally acquiring and outputting a final prediction paper for air quality corresponding to the spatial variation value of the quality prediction value and air quality information.

Images