Spatiotemporal Prediction of PM2.5 Concentrations at Different Time Granularities Using IDW-BLSTM

As air pollution becomes an increasing concern globally, governments, and research institutions have attached great importance to air quality prediction to help give early warnings and prevent the impacts of air pollution. The existing prediction methods for air quality forecasting include deterministic methods, statistical methods, machine learning, and deep learning methods. Deep learning-based prediction methods have attracted much attention these years due to its high performance and powerful modeling capability. However, the majority of the deep learning methods only focus on the prediction of the places where there have monitoring stations, and limited studies have integrated deep learning to predict places without monitoring stations. To address the limitations, this paper proposes a new methodology framework combining a deep learning network, namely, bi-directional long short-term memory (BLSTM) network and the inverse distance weighting (IDW) technique for the spatiotemporal predictions of air pollutants at different time granularities. The BLSTM can effectively capture the long-term temporal mechanism of air pollution. The IDW layer, on the other hand, can consider the spatial correlation of air pollution and interpolate the spatial distribution. A case study is conducted to validate the effectiveness of the proposed methodology. The PM2.5 concentration at Guangdong, China is forecasted. Prediction performances of the LSTM network at hourly, daily, and weekly granularities and over different time spans are presented. Spatial distribution of the predicted PM2.5 concentrations and the prediction errors are analyzed. The experimental results demonstrate that the proposed method can achieve better prediction performance for the PM2.5 concentration compared with other models.