A Weighted Ensemble Approach to Real-Time Prediction of Suspended Particulate Matter

Given the widespread health effects of air pollution, it is imperative to predict air pollution ahead of time. A number of time-series forecasting models have been developed to predict air-pollution. However, an evaluation of individual and ensemble models for real-time air pollution forecasting lacks in the literature. The primary objective of this research is to develop and test individual and ensemble time-series forecasting models for real-time forecasting of air pollution ahead in time. Air pollution data of suspended particulate matter (PM_2.5) over 5-years from Beijing, China was used for model comparisons. The PM_2.5-time-series was split as the first 80% for training and the latter 20% for testing time-series forecasting models. Five individual time-series forecasting models, namely, Multilayer Perceptron (MLP), Convolution Neural Network (CNN), Long-Short Term Memory (LSTM), and Seasonal Autoregressive Integrated Moving Average (SARIMA) were developed. Also, a new weighted ensemble model of these individual models was developed. Among the individual models, results revealed that both during training and test, the CNN performed the best, and this model was followed by the LSTM, MLP, and SARIMA models. Furthermore, the weighted ensemble model performed the best among all models. We highlight the potential of using the weighted ensemble approach for real-time forecasting of suspended particulate matter.