Multi-step ahead hourly forecasting of air quality indices in Australia: Application of an optimal time-varying decomposition-based ensemble deep learning algorithm

Recently, researchers have prioritized the accurate forecasting of the particulate matter (PM) air quality indicators PM2.5 and PM10 in urban and industrial locations due to their importance for the human health. However, accurate short-term forecasting via traditional data mining methods is limited due to the complex nature of these indices on hourly scale. To address this issue, a novel deep learning framework composed of classification and regression tree (CART) feature selection, time-varying filter-based empirical mode decomposition (TVF-EMD), and ensemble deep random vector functional link (DRVFL) scheme to forecast 1-h and 3-h ahead PM2.5 and PM10 in two different zones of Australia. In the first pre-processing phase, CART feature selection optimizes the antecedent information (lag) for each target using the computed important factor. Then, the original signal is decomposed into intrinsic mode functions (IMFs) and residual sub-components by the TVF-EMD to overcome the non-stationarity and complexity. The DRVFL approach is then executed by applying the significant lagged-time components as the optimal input feature via each decomposed sub-component. Finally, all the forecast values obtained using the sub-components are summed up to construct the PM values for each horizon. In addition, the classical RVFL and bidirectional gated recurrent unit neural network (Bi-GRU) models are adopted in the form of hybrid and corresponding standalone frameworks to validate the TVD-EMD-DRVFL. The results demonstrate that TVF-EMD-DRVFL model provides best preciseness with R, RMSE, MAPE, NSE, KGE, IA, U95%, and KGE at t+1 and t+3 to forecast PM2.5 and PM10 for Brisbane and North Parramatta station.