Assessment and prediction of PM₁₀ concentration during haze event in Malaysia using quantile analysis

Haze event in Malaysia occurs typically during the summer monsoon season. Consequently, high atmospheric particles, particularly PM₁₀, were recorded mainly by transboundary air pollution from the neighboring country, affecting human health and the environment. The air pollutants were widely forecasted in various studies previously, especially PM₁₀. However, there was a lack of research conducted specifically during haze events. Therefore, this research aims to develop a reliable modified quantile regression (QR) forecasting model for the next-day (PM₁₀+24), the next-two-day (PM₁₀+48), and the next-three-day (PM₁₀+72) of PM₁₀ levels during a haze event. The development of a PM₁₀ prediction model specifically for haze events play a crucial role in managing and mitigating the impacts of haze on society, making them as essential tools in air quality management and decision-making. Hourly PM₁₀, air quality parameters, and weather parameters datasets at Klang, Melaka, Pasir Gudang, and Petaling Jaya during historical haze events in 1997, 2005, 2013, and 2015 are obtained from the Department of Environment (DOE) Malaysia. The locations were chosen due to their susceptibility to pollution transported from the Sumatra region, being situated on the west coast of peninsular Malaysia Peninsular. The mean value for each year at all location exceeded the Recommended Malaysian Ambient Air Quality Standard (RMAAQG), except for at Pasir Gudang in year 1997 and 2005, where the mean recorded are 47.72 and 46.59 μg/m3, respectively. Three feature selection methods (weight by Relief, weight by correlation, and weight by principal component analysis (PCA) along with quantile regression (QR) and multiple linear regression (MLR) were implemented in this study. The performance of modified QR model was evaluated by using several performance indicators namely Root Mean Squared Error (RMSE), Mean Absolute Error (MAE) and Normalized Absolute Error (NAE). The modified QR model i.e. QR-Relief, QR-PCA and QR-Correlation models were proven as the best method at all locations and prediction time. In Klang, QR-Relief was chosen for PM₁₀+24 (with percent reduction error of approximately 21.5%) and PM₁₀+72 (with percent reduction error of 11.6%) meanwhile for PM₁₀+48 prediction, the QR-Correlation was selected with percent reduction error of 17.4%. QR-PCA was chosen as the best prediction model for all three days predictions in Melaka with error reduced by 2.08%, 0.69%, and 0.88%, for PM₁₀+24, PM₁₀+48, and PM₁₀+72, respectively. In Pasir Gudang, QR-Relief performed the best for all three days predictions with error reduced by 27.6% until 31.1%. In Petaling Jaya, QR-Relief outperformed other models for PM₁₀+24 (with percent error reduction of 16.5%). Meanwhile, QR-Correlation is the model with the best performance for PM₁₀+48 (with error reduced by 10.9%) and PM₁₀+72 (with error reduced by 15.9%) in Petaling Jaya. Feature selection helps identify and include only the most relevant variables in the model which eventually improve the models accuracy. The verification of the models using the unseen dataset from 2019 proved that the model can be deployed in the real-world PM₁₀ data. This proposed model can be used as a tool for early warning alerts to the local authorities to mitigate and plan preventive measures.