Machine learning-enabled estimation and high-resolution forecasting of atmospheric VOCs

Abstract

Atmospheric volatile organic compounds (VOCs) are crucial to reducing air pollution, which has adverse effects on human health. However, VOCs estimation and forecasting has been limited by insufficient observational data and complex interactions with other pollutants. Here, we developed machine learning models to estimate regional VOCs distributions and produce hourly distribution maps for the next 24 h with a 1 km resolution. Combining VOCs observations from monitoring sites, along with meteorological, emission, geographical and other related variables, we successfully employed a space-time LightGBM model to estimate VOCs concentrations from 2019 to 2021 and created a high-resolution uninterrupted VOCs dataset in Shanghai. Using this dataset, we evaluated the forecasting performance of three machine learning models and one deep learning model, finding that LightGBM outperformed other models. The models demonstrated substantial efficacy, with R² values ranging from 0.527 to 0.938 and MAE values between 41.5 and 126 ppb, indicating significant performance in both temporal and spatial scales. With developed models, we provided first high-resolution hourly VOCs prediction maps in Shanghai, providing valuable insights for control strategy formulation in advance. This study also offers a novel tool for VOCs forecasting, with the developed model being adaptable to other regions experiencing high VOCs levels.