Publications

Mao, X; Shen, T; Feng, X (2017). Prediction of hourly ground-level PM2.5 concentrations 3 days in advance using neural networks with satellite data in eastern China. ATMOSPHERIC POLLUTION RESEARCH, 8(6), 1005-1015.

Abstract
This study is an attempt to explore the effectiveness of satellite data in predicting hourly PM2.5 (Respirable particulate matter with aerodynamic diameter below 2.5 mu m) concentrations for a chosen number of forward time steps over eastern China. MODIS (Moderate Resolution Imaging Spectroradiometer) aerosol optical depth (AOD), hourly forecasted meteorological variables, along with respective pollutant predictors from 2013 to 2015 were used as input to a multi-layer perceptron (MLP) type of back-propagation neural network. A novel approach, based on the correlation coefficients between surface PM2.5 and meteorological variables, was employed in selecting the averaging periods for meteorological input. Backward air mass trajectory was combined with AOD so as to explicitly measure the contribution of regional transport. Owing to the spatial variability of the AOD-PM2.5 relationship, each grid cell on the AOD retrievals was assigned to a prediction model trained by the nearest monitoring station from them. The proposed model seems to perform better in southern China. Also predictions incorporating transport predictor tend to have higher rates of detecting PM2.5 exceedance hours. We further introduced a sensitive analysis in Beijing by testing a model with surface PM2.5 input, evaluated versus the AOD input one. The two models were evaluated on the perturbed test set with input values ranging from 10 to 40%. It is found that PM model significantly outperforms the AOD one when the prediction horizon is longer than 36 h. This approach shows potential to be adapted for other regions, but will be most useful for areas without access to sophisticated deterministic models. (C) 2017 Turkish National Committee for Air Pollution Research and Control. Production and hosting by Elsevier B.V. All rights reserved.

DOI:
10.1016/j.apr.2017.04.002

ISSN:
1309-1042