A novel ensemble reinforcement learning gated unit model for daily PM2.5 forecasting

PM2.5 forecasting is an important scientific way to control environmental pollution and keep people healthy. To achieve high-performance PM2.5 forecasting, a new ensemble reinforcement learning gated unit model is presented in this study. The complete framework of this model mainly includes the following steps: In step I, the WPD method is applied to decompose PM2.5 data into 8 sub-series with different frequency types. In step II, the SAE-GRU method is presented to finish the establishment of sub-series forecasting model. Among them, SAE is used to obtain low-latitude features of PM2.5 data, and GRU is applied to finish PM2.5 sub-series forecasting. In step III, Q-learning is used to combine the every PM2.5 sub-series to get the final model prediction results. By comparing and analyzing the final results of all case study, it can be summarized that (1) Q-learning-based ensemble model integrates the subseries with different frequency types perfectly, and results prove that it is better than heuristic algorithm, and (2) the proposed ensemble reinforcement learning gated unit model can get prediction results beyond seventeen alternative models which include three most state-of-the-art models in all cases.