Probabilistic forecasting for air pollutants in Seoul

Abstract

Air pollutants are the known cause of many respiratory and cardiovascular diseases, thus being detrimental influence on human health worldwide. On top of inflicting public health, air pollutants are known to impact the economy and the policy-making process. Therefore, making accurate forecasts of air pollutants is an important matter for all sectors. We conduct probabilistic forecasts instead of point forecasts, as probabilistic forecasts allow us to prepare for events with low probability. We test probabilistic forecasting models based on traditional time series models, namely ARIMA-GARCH against the deep neural network-based time series model DeepAR. As previous research has investigated the correlation between air pollutants and other variables such as weather or traffic volume, we incorporate such covariates to our DeepAR model expecting the covariates to raise model performance. Are results show that DeepAR consistently outputs better performance for probabilistic forecasts in air pollutants compared to ARIMA-GARCH models in most cases, while incorporating covariates to DeepAR further enhances model performance for the accordingly relevant air pollutants.