Prediction of Ambient PM10 Concentration with Artificial Neural Network

This paper presents the design of building economic and flexible artificial neural network (ANN) model for the prediction of PM10 concentration 24-hour-ahead. In stead of relying purely on historical information of the pollutant, the developed model would incorporate the effects from local meteorological conditions and other related pollutants explicitly. The ANN used was a three-layer feed-forward network (TLFN) of the back-propagation type. Computation efficiency was achieved by limiting the size of the input data to six parameters per input set, assuming variation of the predicted hourly PM10 concentration to depend only on meteorological and air quality conditions within the last 72 hours, and separating the model development according to prevailing seasons. Two sets of model for the summer and winter seasons were developed and tested based on one full year of measurements in Macau. Selections of input parameters for models were determined by analyzing correlation coefficients among the hourly concentrations of measured pollutants and seven meteorological parameters. The number of neuron used in the hidden layer for each model was then determined by systematic trials and selecting that with minimum root mean square error. Five and six neurons were determined for the summer and winter models, respectively. Predictions for seven days by the trained models were compared with measurements. Results show that on average half of the predictions achieved accuracy on absolute relative percentage error of less than 50% with the summer model performed slightly better. Further studies on model selection technique are recommended for improvement of prediction accuracy.