Exploring PM_2.5 and PM₁₀ ML forecasting models: a comparative study in the UAE

Particulate Matters PM and PM present a major health and environmental concern in urban regions. This research compares machine learning and time series models, such as Decision Tree (DT), Random Forest (RF), Support Vector Regression (SVR), Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), and Facebook Prophet, for predictions of these matters. Their performances have been evaluated over 1-2 hours, 1 day and 1 week forecasting periods using five years real-life data from six ground stations in Abu Dhabi, UAE. Performance metrics including Mean Absolute Percentage Error (MAPE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and Percent Bias (PBIAS) were applied. Linear SVR was generally the best performing model for PM predictions at all stations with averages of 18.7% and 28.2% MAPE for 1 and 2-hour periods, respectively. However, CNN performed best in forecasting PM for 1-hour horizon, with an average MAPE of 12.6%. For the 2-hour forecast, SVR outperformed other models, with 18.3% MAPE. Facebook Prophet consistently outperformed others for both PM and PM with 21.8% and 13.4% MAPE for 1-day and 21.3% and 13.8% MAPE for 1-week, respectively. These best performing models yielded similar RMSE, MAE, and PBIAS values for both PM and PM.