A Lag-FLSTM deep learning network based on Bayesian Optimization for multi-sequential-variant PM2.5 prediction

To better support the prevention of air pollutions for sustainable cities, researchers have studied different methods to forecast air pollutant concentrations. Existing methods have gone through the development from deterministic methods, statistical methods, to machine learning and deep learning methods. The latest direction lies in Long Short-Term Memory (LSTM) based methods. They are a special kind of deep learning network, and can not only well model non-linear real-world problems, but also consider the impact of long-historical values. These methods have achieved state-of-the-art performance in air quality predictions, but some gaps have not been well addressed, especially the overlook on the multi-sequential-variants, and the lack of efficient parameter optimization in the deep learning models. To this end, this study proposes a Lag-FLSTM (Lag layer-LSTM-Fully Connected network) model based on Bayesian Optimization (BO) for multivariant air quality prediction. A case study in the U.S. is conducted to test the method. The results showed that Lag-FLSTM has at least 23.86 % lower RMSE than other methods. The contributions of this study are that we not only developed a deep learning method that can automatically optimize the model parameters but also studied how different metrological features and other pollutants affect the prediction of PM2.5 concentrations.