Stacked Selective Ensemble for PM2.5 Forecast

Abstract

In this paper, we propose a new stacked selective ensemble-backed predictor (SSEP) to forecast the concentration of PM2.5 based on the impact of measurements of the known air pollutants and meteorological data on the unknown PM2.5 concentration over the following 48 h. It was found that a single learner cannot validly uncover and model the relationship between the future PM2.5 concentration and the current and historical meteorological and pollutant data, mainly because any individual learner has limitations, especially facing to highly complex and ever-changing environmental problems, such as PM2.5 prediction. Ensemble methods, which are widely acknowledged to yield strong generalization ability by boosting weak learners, are used in this paper to solve the aforesaid challenge. Our solution, aligned with an analysis of influencing factors on the future PM2.5 concentration, generates multiple component learners for aggregation by introducing three types of diversities. Then, we adopt a pruning method to remove the negative component learners in each diverse type according to dynamic thresholds, which are determined by jointly considering the performance of each individual learner and the correlations between each pair of learners. A stacking technique is finally applied to the selected positive component learners to forecast the PM2.5 concentration in the future. Thorough experiments demonstrate the superiority of our proposed SSEP in contrast to relevant state-of-the-art models when applied to PM2.5 prediction.