Dempster-Shafer ensemble learning framework for air pollution nowcasting

Deep-learning has emerged as a powerful approach to significantly improve forecast accuracy for air quality estimation. Several models have been developed, demonstrating their own merits in some scenarios and for certain pollutants. In nowcasting, the prediction of air pollution over a small time period essentially demands accurate and reliable estimates, especially in the event cases. From these, selecting the most suitable model to achieve the required forecast performance remains challenging. This paper presents an ensemble framework based on the Dempster-Shafer theory for data fusion to identify the most accurate and reliable forecasts of air pollution obtained from multiple deep neural network models. Our framework is evaluated against three popular machine learning methods, namely, LightGBM, Random Forest, and XGBoost. Experiments are conducted on two horizons: 6-hour and 12-hour predictions using real-world air quality data collected from state-run monitoring stations and low-cost wireless sensor networks.