Recurrent Neural Network and random forest for analysis and accurate forecast of atmospheric pollutants: A case study in Hangzhou, China

Abstract

Hangzhou, one the most prosperous cities in China, suffers from severe atmospheric quality degradation in recent years. For the good of nine million local citizens and incoming Asian Games, Recurrent Neural Network (RNN) and Random Forest are used to analyze the air pollution in Hangzhou. Compared with the traditional atmospheric models, machine learning models are faster, more accurate and less costly in simulating all the pollutants without using the pollution inventory. The Feature Importance (FI) generated by Random Forest reveals the complicated relationships among air pollutants and meteorology. Carbon monoxide (CO) plays an important role in shaping ground-level ozone, nitrogen dioxide (NO2) and particulate matters (PM) in the atmospheric environment. Dew-point deficit plays a more important role than relative humidity in shaping air pollutants. Urban heat island effect is not obvious for the air pollutants from non-point source. Furthermore, a WRF/RNN-based method to forecast air pollutants, including SO2, NO2, CO, PM2.5, PM10 and O3, in the future 24 h is proposed and a RNN-based method to estimate regional transport rate of air pollutants and reversely identify air pollution emission sources is introduced. At last, policy assessment is made to better regulate the air pollution in Hangzhou, prior to the 2022 Asian Games.