Spatial-Temporal Prediction of PM2.5 Pollutants Using Deep Recurrent Networks: A Case Study of Tehran

In recent years, air pollution has become one of the most important environmental challenges in large and industrial cities such as Tehran. High concentration of particulate matter with a diameter of less than 2.5 μm (PM2.5), which is known as the main cause of pollution in Tehran, is associated with irreversible effects on human health. Providing spatial-temporal model with high accuracy and speed for forecasting, is an effective way to protect public health against the increase of harmful air pollutants. The rapid growth of computing technologies and the availability of air quality data have provided researchers with the opportunity to provide sophisticated models in the context of machine learning, especially in deep learning to predict the concentrations of various air pollutants. In this study, with the aim of predicting PM2.5 concentrations at different time intervals, a new spatio-temporal deep learning model based on gated recurrent units (GRU) is presented which maintains and extracts temporal and spatial dependencies in the time series of air pollution datasets. The proposed model has been compared with support vector machine regression (SVR) and long-term memory (LSTM) methods as competitive approaches. The data used in this study include the hourly concentration of PM2.5 and meteorological parameters recorded by 13 air pollution monitoring stations and 3 synoptic meteorological stations in Tehran in the period of December, 2016 to February, 2019, respectively. The model presented in this paper with the RMSE of 7.97 μg/m3 and MAE of 5.35 μg/m3 has the best result for predicting air contamination compared to other methods. This model can determine 80% (R2=80) of PM2.5 concentration changes and predict contamination level. The proposed model also proves that it can be used effectively to predict and control air pollution by extracting temporal properties, simultaneous forecasting for all stations and considering spatial correlations.