Enhancing Short-Term PM2.5 Prediction in Shah Alam using Wrapper Feature Selection and Machine Learning Techniques

Background: Air pollution poses significant health and environmental risks in Malaysia, particularly from fine particulate matter (PM2.5), which can cause cardiovascular and respiratory problems. PM2.5 originates from both natural and human activities and is a key indicator of air quality. In Malaysia, authorities typically require at least three days to respond to hazardous pollution levels. This study aims to forecast PM2.5 levels over the next three days in Shah Alam, a city with high air pollution rates. Methods: To extract the most influential variables affecting PM2.5 levels among the meteorological and the gas pollutant variables, a wrapper feature selection technique was employed. The selected subset of features was used to train multiple machine learning models, which were compared using various metrics. To improve the predictions made by these models, two ensemble learning techniques were utilized to combine them. Results: As a finding, the most effective model was the Blending Regressor, an ensemble model that combines predictions and achieves an R2 of 0.70 and a Root Mean Squared Error of 6.32. Conclusion: The study identified that the Blending Regressor with forward selection achieved the best performance in predicting PM2.5 levels, highlighting the importance of gas concentration variables. To improve model accuracy, incorporating advanced machine learning techniques and alternative feature selection methods is recommended.