Macau Ambient Air Pollutants Forecasting By Pollutant Time Series And Meteorological Data Using Machine Learning Methods

Air pollution in Macau has become a serious problem following the Pearl River Delta’s (PRD) rapid industrialization that began in the 1990s. While there has been continual improvement in recent years, harmful air pollutant concentration levels are still common, impacting Macau residents' health and creating long-term medical costs to local society. With this in mind, Macau needs an air quality forecast system that accurately predicts pollutant concentration and an early alert system instead of only daily real-time reminders.
Some scholars have previously carried out studies to develop an air quality forecast for Macau by successfully using statistical models. Therefore, pursuant to the outcomes of previous studies, this dissertation aims to build upon research results and explore further possibilities of building a better ML air quality forecast model based on the time series of air pollutants concentration and meteorological data.
Four different state-of-the-art ML algorithms were used to create predictive models to forecast PM2.5, PM10, and carbon monoxide (CO) concentrations for the next 24 and 48- hour. These were Support Vector Machine (SVM), Artificial Neural Networks (ANN), Random Forest (RF), and Extreme Gradient Boosting (XGBoost). In addition, Multiple Linear Regression MLR, a standard ML model, was used for this dissertation as a baseline reference for performance comparison.
The daily measurements of air quality data in Macau from 2016 to 2021 were collected for this dissertation. The 2020 and 2021 datasets were used for model testing while the four- year data prior to 2020 and 2021 were used to build and train the ML models.
The results showed that SVM, ANN, RF, and XGBoost were able to provide a very good performance in building up a 24-hour forecast with higher R2 and lower RMSE, MAE, and BIAS. Meanwhile, all ML models in 48-hour forecasting performance were satisfactory enough to be accepted as a two-day continuous forecast even if the R2 value was lower than the 24-hour forecast. The 48-hour forecasting model could be further improved by proper feature selection based on the 24-hour dataset, using the SHAP value test, and the adjusted R2 value of the 48-hour forecasting model.