APPLICATION OF ARTIFICIAL NEURAL NETWORKS AND REGRESSION MODELS IN THE PREDICTION OF DAILY MAXIMUM PM10 CONCENTRATION IN DUZCE, TUKEY

Increasing levels of atmospheric particulate matter are known to adversely affect human health. Therefore, air quality predictions may provide important information in order to take actions for the public before the pollution happens. In this study, we presented artificial neural network (ANN), stepwise regression (SR) and multiple linear regression (MLR) models to forecast maximum daily PM10, concentrations one day ahead in Duzce, Turkey. Particularly, a special emphasis was put on the prediction of particulate levels during winter episodes. Inputs to the models include lagged values of maximum, minimum and standard deviations of PM10, concentrations, and some meteorological factors, which are all on daily basis. The output is the expected maximum concentration of PM10 in mu g.m(-3) for the following day. The data sets used in training and testing stages covered the daily averaged values of these variables for the period of 2011-2013. The results showed that selected inputs based on stepwise regression approach and use of cascading-training in multi-layer perceptron ANN (ANN-MLP) appeared to be promising with R-2 up to 0.69 and index-of-agreement up to 0.79. It is concluded that local monitoring systems associated with ANN model predictions may be a sound way to develop embedded online systems for public health.