Spatial-temporal characterization of air pollutants using a hybrid deep learning/Kriging model incorporated with a weather normalization technique

A scarce distribution of the PM_2.5 chemical compositions monitors reduces the applicability of scientific information for policymakers to assess the effectiveness of air pollution control strategies. There is an urgent need for a spatial-temporal prediction model for characterizing PM_2.5 chemical compositions to assess exposure risks and develop effective air pollutants reduction strategies. In this study, the spatial-temporal variations of NO₃⁻ and SO₄²⁻ were characterized using a hybrid multi-step ahead neural network (MSA-NN)/Kriging model in the urban areas with limited PM_2.5 constituents monitoring stations. A meteorological normalization technique was further applied to develop a de-weather model to investigate temporal variations of air pollutants during the level 3 COVID-19 alert in central Taiwan. The MSA-NN algorithm could predict 94% and 91% of NO₃⁻ and SO₄²⁻, respectively, at the t+1-time horizon predictions. Based on the predicted results using the present de-weather model, the reduction in primary emissions attributed to the impact of COVID-19 during the level 3 alert was found to dominate the temporal air pollutant concentrations in central Taiwan. The present model could provide applicable and accurate high resolution of spatial-temporal NO₃⁻ and SO₄²⁻ datasets in an area with limited PM_2.5 chemical composition measurement. The present model could also be potentially applied to facilitate hotspot identification and human exposure assessment. The present Artificial Neural Network-based de-weather model is applicable to predict meteorological normalized time series air pollutant concentrations, which could be used to verify the effects of the meteorological parameters and primary emissions on the variations in air quality during the implementation of a specific air quality control strategy or changes in anthropogenic activities.