Artificial Intelligence for Ground-level Ozone Concentration Forecasting Using Data From the Ground Stations of the Abu Dhabi Environment Agency

Tropospheric ozone
$(\mathrm{O}_{3})$
is a secondary pollutant generated from the photochemical reactions of two pollutants: nitrogen oxides and volatile organic compounds.
$\mathrm{O}_{3}$
higher concentration above the earth's surface harms human health and ecosystems, which urges the need to build a robust model that accurately forecasts pollutant concentration to support decision-makers in mitigating its adverse effects. In this study, we compare the performance of four state-of-the-art deep learning models for temporal data to forecast pollutant future concentration using five air pollution stations that exhibit different environmental assessment points. Overall, the LSTM and Transformer-based models outperform other models. The Transformer model reported a lowest RMSE of 0.25 in South Habshan. At the same time, LSTM reported the best performance for the Ruwais station with RMSE 0.47. Incorporating deep learning techniques can significantly enhance the prediction of ozone concentration. We also have observed that the temporal characteristic of the pollutant can impact the model's performance. The Transformer-based model excels when the pollutant sequence has great diversity. In contrast, LSTM stands out with a lower variation sequence.