A neural network model for three-hours-ahead prediction of ozone concentration in the urban area of Palermo

摘要

The purpose of this study is to use a recurrent neural network (Jordan model) to forecast ozone concentrations (O_3) with a short lead-time (1-3h) in the lower atmosphere. The network has been trained using a time series that was recorded between January 1st 2003 to December 31st 2003 and at two monitoring stations in Palermo (Italy). Each input pattern is composed of twelve (hourly) values: wind direction and intensity, barometric pressure, and ambient temperature; respectively gathered at the meteorological stations in Bellolampo, Boccadifalco and Castelnuovo. Ozone predictions are notoriously complex when using either deterministic or stochastic models which explains why this model was developed using a Neural Network. Neural Networks possess the ability to learn about nonlinear relationships between the variables used. The model developed is a potential tool for the predictions air quality parameters and it is superior to the traditional stochastic model.