Regional air quality forecasting using spatiotemporal deep learning

Accelerated urbanization and industrialization have led to poor air quality, which threatens human health with various lung ailments. Monitoring, modeling, and forecasting air quality would be a prudent way to promote awareness and defend human beings from the adversities of air pollution. The air quality of a region is monitored through various air quality monitoring stations built in and around it. The air quality data collected from these stations are highly dynamic, nonlinear, and hold intensely stochastic spatiotemporal correlations in them. Deep learning algorithms that are capable of extracting a higher level of abstraction in data can efficiently capture the spatiotemporal features in it. In this paper, we propose a hierarchical deep learning model named DL-Air that embodies three components for air quality forecasting. The first component, the encoder, encodes all spatial relations in the data. The second component, STAA-LSTM, a proposed variant of LSTM, identifies all temporal relations and the level of association between the identified spatiotemporal relation and the forecast. Also, the STAA-LSTM predicts the future spatiotemporal relations in the latent space. The third component, decoder suitably decodes these relations to obtain the actual forecast. The proposed framework was extensively evaluated for forecasting the real-world air quality data of Delhi. DL-Air shows better performance with around 30% reduced RMSE and MAE, 37% reduced AAD, 11% improved R-2 and 8% improved accuracy in AQI category prediction than the best performing baseline approaches. Also, the predictive performance of DL-Air is found to be consistent across all seasons in Delhi. (C) 2020 Elsevier Ltd. All rights reserved.

Automated summary

Accelerated urbanization and industrialization have worsened air quality, but monitoring, modeling, and forecasting can help raise awareness and protect human health. Deep learning algorithms can efficiently capture the spatiotemporal features of air quality data.