Air Quality Prediction Using a Deep Neural Network Model

A deep neural network (DNN) model of multi-layer perceptron with 3 or 4 hidden layers is developed to predict the air qualities. The DNN model takes the past 3 days of the hourly concentration measurements of the pollutants (CO, SO2, NO2, O-3, PM10, PM2.5) and the meteorology data (wind speed, wind direction, air temperature, air humidity), and then predicts the hourly concentration of the pollutants for the next 24 hours. The DNN model was compared against the observations from all nationwide air quality monitoring stations which includes 115 sites in 7 metropolitan cities in South Korea. The index of agreement (IOA) was found to be 0.7 similar to 0.8, based upon the 6,505 comparison data sets from January 1, 2017 to September 30, 2017. In the unit of air quality grade, which can be evaluated from the pollutant concentration level, 60%similar to 80% cases of the DNN predictions agree with those of the observations. For the region-wide PM10 grade, the DNN predicts exactly the 75%similar to 85% cases of the observations, which is in about the same accuracy range of the numerical air quality models of the current operative use. Yet, for the region-wide PM(2.5 )grade, the cases of the accurate predictions of DNN is about twice of those of the numerical model. In the metropolitan Gwangju, for an example, the DNN predicts exactly the 211 next days of the PM2.5 grade, while the numerical model forecasts just 120 days correctly.