Multi hours ahead prediction of surface ozone gas concentration: Robust artificial intelligence approach

Abstract

Forecasting the Ozone concentration is a substantial process in many important environmental issues such air pollution management, risk assessment, public health, and global warming. Early and accurate prediction of Ozone is very significant for efficient monitoring of pollution towards constituting advanced early alarm systems. In this study, wavelet transform (WT) approach was applied to handle input variables prior to introduce them to artificial neural network (ANN). This approach attempts to remove the noise impacts which decreases the accuracy of simulation processes. Additionally, to improve ANN model performance, the selection of suitable type of transfer function with effective input combinations was thoroughly investigated before introducing the WT approach. The hybrid model (W-ANN) was also compared to the classical ANN in the prediction of 1 h ahead Ozone concentrations to insure the significance of this study. Based on statistical measures during all phases (i.e. training, validation, and testing), the W-ANN outperformed ANN model. The W-ANN model yielded fewer values of root mean square error (0.9313 ppm) and mean absolute error (0.6531 ppm) while, ANN model produced higher values of root mean square error (2.902 ppm) and mean absolute error (1.991 ppm) during the testing set. The quantity of predicted values that have absolute relative error between 1 and 3% represented 18% and 36.16% of all data points for each approach, ANN and W-ANN respectively. Additionally, the hybrid W-ANN managed successfully to forecast multiple time scales (2-, 3-, 4 and 5- ahead hour Ozone concentrations) and reported very fewer errors in comparison with modeling of one hour ahead Ozone gas concentration.