Deep learning based spatiotemporal modelling for air quality forecasting

Abstract

Spatiotemporal systems hold data related to both spatial and temporal scales. Unlike relational data, the spatiotemporal data collected from these systems introduces significant challenges for knowledge extraction. Air quality monitoring is a real-time spatiotemporal system. Forecasting the air quality values collected by this system is beneficial to environmental conservation. Accurate spatiotemporal air quality forecasting is a challenging task as the regional air quality data collected from various monitoring stations dispersed across a region contains intricate spatiotemporal correlations in them. The relationship within the air quality values captured by a monitoring station is referred to as intra-dependency in air quality data. While interdependency refers to the relationship between the air quality values captured by two neighboring monitoring stations. Both Intra-dependency and Inter-dependency in the air quality data must be assessed for accurate forecasting. Additionally, the spatiotemporal characteristics of air quality data have non-linear and stochastic relationships with their forecast values, which must be considered for accurate air quality forecasting. Numerical models that can forecast regional air quality values require complex computations. With large amounts of data available, a data-driven approach is a viable option. Statistical models in data-driven approaches overlook spatiotemporal features, making them less suitable for spatiotemporal modeling of air quality data for accurate forecasting. Deep learning is an artificial intelligence algorithm that uses multiple layers of networks to extract layers of abstraction from data. Deep learning algorithms has proved its excellence for extracting spatiotemporal features in several applications such as wind speed forecasting, medical image analysis, epidemiology, taxi-demand prediction, and so on.