An ensemble multi-step-ahead forecasting system for fine particulate matter in urban areas. (1st August 2020)

• Record Type: Journal Article
• Authors: Ahani, Ida Kalate; Salari, Majid; Shadman, Alireza
• Is Part Of: Journal of cleaner production. Volume 263(2020)
• Publication Date: 2020-08-01
• Subjects: Multi-step-ahead -- Fine particulate matter -- Long-term forecasting -- Neural network -- Least square support vector regression -- Long short-term memory; Factory and trade waste -- Management -- Periodicals; Manufactures -- Environmental aspects -- Periodicals; Déchets industriels -- Gestion -- Périodiques; Usines -- Aspect de l'environnement -- Périodiques; 628.5
• DOI: 10.1016/j.jclepro.2020.120983
• Languages: English
• ISSNs: 0959-6526

Abstract

Abstract: In recent years, growing air pollution has become a significant issue due to its detrimental effects on the environment and different living organisms. Providing accurate and reliable forecasts of air quality over a long future horizon is an effective way to mitigate health risks. In this paper, the problem of urban PM2.5 forecasts for several days ahead is considered. An ensemble multi-step-ahead forecasting system is introduced for this problem, which combines different multi-step-ahead strategies (including single-output and multi-output approaches). The proposed hybrid framework consists of three parts. In the first part, the Ensemble Empirical Mode Decomposition (EEMD) technique is combined with a prediction tool and multi-step-ahead strategies. Boosting idea is considered in the second part of the algorithm. Finally, the stacked ensemble of boosted hybrid structures is developed to provide the final multi-step-ahead forecasts. Least Square Support Vector Regression (LSSVR), and Long Short-Term Memory neural network (LSTM) are employed as the prediction tools in the proposed hybrid framework. Through real PM2.5 data examples from Mashhad, Iran, the proposed ensemble model is investigated for 1-day-ahead to 10-days-ahead. The results reveal the effectiveness of the ensemble model in comparison with the multi-step-ahead strategies in all time-steps. The proposed model with LSSVR prediction tool shows the smallest mean RMSE, MAE, and MAPE values of 7.810, 5.562, and 18.104% over all time-steps, and RMSE improvement rates of more than 35% compared to simply combining different multi-step-ahead strategies with LSSVR approach.