Deep learning models for air quality forecasting based on spatiotemporal characteristics of data

Abstract

The distribution of air-borne pollutants is governed by complex fluid dynamics processes involving convection and diffusion. The process is further affected by the characteristics of emission sources, meteorological parameters, socioeconomic factors, and land use patterns. Compared to deterministic and probabilistic air quality forecasting methods, data driven modeling of air quality parameters can address the large degree of freedom in air quality influencing parameters as well as offer interpretability and understanding of air pollutants' distribution at an increased spatial and temporal resolutions. This study focuses on the citywide prediction of air quality index (AQI) based on observations of pollutant concentrations, meteorological parameters, and spatiotemporal data. The study area includes Ansan city in South Korea, which has been observed as a hotspot for high concentrations of particulate matter. The air quality and meteorological were collected from 16 monitoring stations located in Ansan city. A detailed spatiotemporal analysis was performed to investigate the correlation between AQI records at the air quality monitoring stations. Based on strong spatiotemporal correlations observed between stations, several deep learning (DL) models were proposed, and their performance was investigated for different scenarios. It was observed that the selection of appropriate DL models should be based on (1) understanding of the underlying fluid dynamics process that control pollutant distribution and (2) spatiotemporal characteristics of data. Additionally, the complexity of DL models does not always guarantee the accuracy of the forecasts, and simple models can give good performance if the predictors are selected carefully to reflect the underlying physical process.