Ozone prediction based on neural networks and Gaussian processes

The urban environment in Slovenia is confronted with the air pollution problem of harmfully high ozone concentrations. In the last two decades the automatic ozone measuring network was extended and now covers regions where the highest values are expected. Due to topographical and climatological conditions and the presence of extensive urban environments, the most critical locations are the ones in the western part of Slovenia. In the city of Nova Gorica a modern automatic urban air pollution measuring station was installed. Measurements at this station clearly showed that ozone is a considerable pollutant there, especially in the summer time. In this work a perceptron neural-network-based model and a Gaussian-process-based model for ozone concentration forecasting for the city of Nova Gorica was developed and evaluated. The methods of feature determination and pattern selection for the model training process are delineated. The shortcomings of the models and possibilities for improvements are discussed with respect to evaluation of the effectiveness of the methods.