A new multi-data-driven spatiotemporal PM2.5 forecasting model based on an ensemble graph reinforcement learning convolutional network

Spatiotemporal PM2.5 forecasting technology plays an important role in urban traffic environment management and planning. In order to establish a satisfactory high-precision PM2.5 prediction model, a new multidata-driven spatiotemporal PM2.5 forecasting model is proposed in this paper. The overall modelling framework consists of three main parts. In part I, the graph convolutional network uses an adjacency matrix to effectively aggregate spatiotemporal pollutant data from different nodes and extract the most valuable feature information for target point modeling from the original data. In part II, the extracted feature information is used as the input of the gated recursive unit and the long short-term memory network to construct the prediction model. In part III, the Q-learning algorithm builds the best ensemble PM2.5 forecasting model by analyzing the processing ability and analysis ability of different predictors. Based on the analysis of multiple cases, the following conclusions can be drawn: (1) Graphic convolutional networks can effectively analyze the spatiotemporal correlation of PM2.5 data and achieve better performance than traditional convolutional neural networks. (2) Q-learning can adaptively optimize the ensemble weight coefficient and achieve better results than the traditional optimization algorithm. (3) The proposed GCN-LSTM-GRU-Q model can achieve better results than the 24 benchmark models. •A New Multi-data Driven Spatiotemporal PM2.5 Forecasting Model is proposed.•GCN extracts the spatialtemporal features of PM2.5 data from different sites.•Q-learning is used as the ensemble method.•The proposed model is compared with twenty-four mainstream forecasting models.