A novel hybrid ensemble model for hourly PM2.5 forecasting using multiple neural networks: a case study in China

• DOI: 10.1007/s11869-020-00895-7
• Publication Date: 2020-08-01T10:02:22Z
• Authors: Hui Liu; Shuqin Dong
• Keywords: Robustness; Multilayer perceptron; Ensemble forecasting; Ensemble Learning

High concentration PM2.5 may cause serious damage to human health. Accurate PM2.5 concentration forecasting can provide the public with timely and effective PM2.5 pollution warning information. In the current mainstream studies, most existing air pollution forecasting models use only one predictor, of which the accuracy and stability can be further improved. In this study, a novel hybrid ensemble model with three deep learning predictors is proposed for hourly PM2.5 concentration forecasting. In the proposed model, the complementary ensemble empirical mode decomposition (CEEMD) is used to extract the features in the PM2.5 data series to reduce its complexity. Three deep neural networks are used as predictors for data forecasting, including deep belief network (DBN), long short-term memory network (LSTM), and multilayer perceptron (MLP). Each predictor is given a weight, and the imperial competition algorithm (ICA) is used to optimize weights to obtain the best forecasting result. Two groups of PM2.5 concentration data from Shanghai are used to validate the model. The experimental results show that the proposed model has high accuracy and robustness, and can outperform all comparison models.