A hybrid deep learning network for forecasting air pollutant concentrations

Air pollution has become a serious problem. Thus, this study formulated a method based on a multi-input multi-output, hybrid, deep neural network for air quality prediction. In our method, the Pearson correlation is used to identify which predictors ought to be included in the model. The network architecture features dilated convolution and a bidirectional gated recurrent unit for the model to learn relationships (particularly temporal relationships) in time-series data; it also features an attention mechanism. Our method outperformed its counterparts in experiments on five real-world data sets in predicting PM2.5 and PM10 concentrations.