Multi-step ahead forecasting of regional air quality using spatial-temporal deep neural networks: A case study of Huaihai Economic Zone

Abstract

Air pollution has endangered both ecological environment and human health. Long-term prediction of air quality index (AQI) is an effective approach to early warning of, and prompt response to, pollution events to support cleaner industrial production. However, existing approaches to forecasting long-term air quality need further improvement. In this paper, we proposed a novel spatial-temporal deep learning algorithm based on bidirectional gated recurrent unit integrated with attention mechanism (BiAGRU), for more accurate air quality forecasting. The historical air quality measurements and meteorological monitoring data were constructed as a spatial-temporal matrix suitable for model input. The performance of the proposed BiAGRU model was evaluated by a series of metrics. The RMSE , MAE , R 2 and Fractional Bias ( FB ) values of the proposed BiAGRU model are 31.10, 23.06, 0.60, and 0.015, respectively, for 24 h multi-step ahead prediction assignments using Huaihai Economic Zone dataset. Quantitative comparison between models indicates the developed BiAGRU model outperformed various traditional machine learning algorithms and advanced deep neural network methods in term of lower error bias and higher accuracy and regression scores. This work is of importance to strengthen regional prevention and control of air pollution. • A novel artificial intelligence methodology for multi-step ahead forecasting and analysis of air quality. • Inclusion of spatial information improves regional air quality forecasting. • Model performance is evaluated against comprehensive metrics. • The novel spatial-temporal BiAGRU model outperformed several state-of-the-art algorithms.