Artificial neural networks based modelling of carbon monoxide: Effects of spatial parameters

Abstract

Web systems for air quality information and pollution modeling provide access to daily measured and predicted valuable information for the internet users. There are just a few examples around the world in this area. A recently developed web system at http://airpol.fatih.edu.tr is one such application. It provides information about air quality prediction for Istanbul, Turkey in a user friendly interface. New studies have been performed to increase the prediction efficiencies of the system and presented in this work.

This paper presents a study on the prediction of carbon monoxide (CO) levels using Artificial Neural Networks (ANN). ANN models have been applied to the prediction of 3 day CO levels into future. The observed and predicted values were compared to determine the performance of the ANN models. The experimental results shows that spatial parameters (Universal Transverse Mercator coordinates) generally produce better forecasting (30% error) than ordinary, non-spatial parameters, although there are some cases where spatial parameters yield lower prediction accuracy. The experiments reveal that models with spatial input variables deserve further study and better models could be developed with higher accuracy.