Forecasting PM2.5 and Tracking Spatial Influence Patterns of Traffic Using Interpretable Deep Learning

Abstract

Air pollution is a growing worldwide problem. Accurate prediction of PM2.5 concentration has a vital role to reduce the dramatic toll of air pollution on health. Due to the non-linearity and complexity of air pollution process and the influence of multiple factors, such as meteorological conditions, human activities and other chemical components, traditional pollution-related models have challenges in dealing with PM2.5 modeling. Based on atmospheric domain knowledge, we proposed a novel and interpretable deep learning model (iDeepAir) to predict hourly PM2.5 concentration by incorporating traffic data, meteorological data and air quality data. We designed feature interaction module and temporal interaction module to simulate pollution chemical reaction process and temporal accumulated process respectively, which makes the model has better understood and improves prediction accuracy of PM2.5 concentration. Compared to the best comparison model, mean absolute error (MAE) and rooted mean squared error (RMSE) were improved by 20.1% and 14.4% in 24h respectively. Furthermore, with the embedded Layerwise Relevance Propagation (LRP) algorithm, iDeepAir allows us to observe the spatial influence patterns of regional traffic emissions in a high-resolution way and evaluate the impact of traffic emissions on PM2.5 formation. Taking Shanghai as an example, we discover that although there are serious traffic emissions in some areas of Shanghai, they do not always directly aggravate air pollution, which is also affected by local buildings, meteorological conditions, and other human activities. These results show the spatial interpretability of our model and provide a quantitive decision-making basis for the government to control air pollution.