A hybrid multi-resolution multi-objective ensemble model and its application for forecasting of daily PM2.5 concentrations

Abstract

PM2.5 concentrations forecasting can provide early air pollution warning information for the public in advance. In this study, a novel multi-resolution ensemble model for multi-step PM2.5 concentrations forecasting is proposed. This model utilizes the high resolution (1-h) and low resolution (1-day) data as the input, and outputs low resolution PM2.5 concentrations forecasting data. For the high resolution data, real-time wavelet packet decomposition (WPD) is applied to generate sub-layers, the features within the high resolution sublayers are extracted by stacked auto-encoder (SAE), and the extracted features are fed into the bidirectional long short term memory (BiLSTM) to generate PM2.5 concentrations forecasting results. For the low resolution data, the forecasting results are obtained by the real-time WPD and BiLSTM. The forecasting results obtained by the high and low resolution data are combined by the non-dominated sorting genetic algorithm (NSGA-II) algorithm to output the deterministic forecasting results. The bivariate kernel density estimation (BKDE) algorithm is applied to describe the heteroscedasticity and non-Gaussian characteristics of the deterministic forecasting residuals and produce probabilistic forecasting results. Four real air pollutant data are utilized to validate the proposed model. The experimental results show the proposed model has better forecasting performances than the benchmark models.