PM10 CONCENTRATION PREDICTION FOR AREAS WITH NO UPDATING MONITORING SYSTEM USING AUTO – REGRESSIVEGROUP METHOD OF DATA HANDLING NEURAL NETWORK

Introduction: Predicting PM10 concentration as a significant risk factor for anumber of pollution related diseases has been recently inevitable task for areas with high population density particularly for areas with no updating monitoring systems. This study aims to illustrate how PM10 concentration level can be predicted by the prior information of the air pollutants and the meteorologicalfactors in urban areas. Materials and methods: The data we used are measured from four monitoringstations in the city of Tehran between January 2012 and December 2014. We use the Auto-regressive group method of data handling (AR - GMDH) neuralnetwork approach which employees the prior stationary time series data setting. Results: Our results demonstrate that PM10 concentration level for a specific dayis more likely to be predictable by sulfur dioxide (SO2) and nitrogen dioxide (NO2) than the carbon monoxide (CO) concentrations, and also show thatPM10 concentration is positively associated with precipitation and wind speedand with high temperature. The accuracy of the predicted values of the PM10 concentration is evaluated by inspecting the coefficient of determination, meansquared error, the square root of mean squared error, mean absolute deviation, and index of agreement. Conclusions: The AR - GMDH algorithm can be proposedin comparison with the chemical and physical approaches due to its accuracyand simplicity, and its cost efficiency.