Suyeon HWANG, Kyung Hee University, Korea (Republic of)
Jinmu CHOI, Kyung Hee University, Korea (Republic of)
Since the fine dust forecast model used by the Korea Meteorological Administration utilizes the fine dust concentration of air pollutant monitoring stations, there is a problem that the accuracy of prediction decreases when missing values occur due to defects in the station. Recently, Neural network is emerging as a new way to increase the accuracy of predicting the concentration of fine dust. However, existing studies have not considered missing values and spatial distribution of stations when making models. Therefore, this study proposed a fine dust concentration prediction model that can consider both concentration data including missing values and the spatial distribution of stations. Long Short Term Memory, a neural network model used in this study, solves the long term dependence problem of Recurrent Neural Network.
First, along with linear interpolation and spatial interpolation, the newly proposed combination interpolation in the study was compared as a missing treatment methods. The model was produced through the selected missing processing method, and a high-accuracy model was selected by combining weather data, concentration data of the monitoring station to be predicted, and concentration data of the neighbor monitoring station selected using the Queen proximity method. The input dataset is a form of predicting the concentration of the next time with data for 24 hours.
As a result of comparing missing methods, the accuracy of the Combination interpolation method was the highest. When making a model using this, it showed higher predictive accuracy to use the concentration of neighbor stations as input data so that spatial distribution between stations as well as weather data can be considered. The model proposed through this study can be used as an efficient prediction method of fine dust concentration that can reflect the shape of the missing concentration data and the spatial distribution of the monitoring station.
Mots clés : Fine dust|PM10|Spatial Interpolation|Neural Network|LSTM
A103494SH