A neural network forecasting system for daily air quality index in Macau

Using the data recorded at the Northern Zone monitoring air-quality monitoring station at Macau during April to June of 1999 as the training and testing data set, a three-layered feed-forward neural network is developed to predict the one-day ahead daily air quality index (AQI) at that area. Various input selection ranging from using the past three day measurements to the past twelve day measurements are tested against different number of hidden layer neurons. It is found experimentally that using only the past three days values as input with 8 neurons in the hidden layer give the best testing results. This setting is adopted as the basic setting of the developed model. It is then applied to forecast the sub-index of each measured pollutant and the predicted AQI is set as the maximum of the forecasted sub-index. The forecasted AQIs are compared with the actual measurements of July 1999. It is found that the mean absolute percentage error is about 16% and 97% of the forecasts give the correct main contribution pollutant. These promising results indicate that an efficient AQI-prediction system for Macau could be developed based on the application of neural network.