Short-Term Air Pollution Forecasting Using Embeddings in Neural Networks

Abstract

Air quality is a highly relevant issue for any developed economy. The high incidence of pollution levels and their impact on human health has attracted the attention of the machine-learning scientific community. We present a study using several machine-learning methods to forecast NO2 concentration using historical pollution data and meteorological variables and apply them to the city of Erfurt, Germany. We propose modelling the time dependency using embedding variables, which enable the model to learn the implicit behaviour of traffic and offers the possibility to elaborate on local events. In addition, the model uses seven meteorological features to forecast the NO2 concentration for the next hours. The forecasting model also uses the seasonality of the pollution levels. Our experimental study shows that promising forecasts can be achieved, especially for holidays and similar occasions which lead to shifts in usual seasonality patterns. While the MAE values of the compared models range from 4.3 to 15, our model achieves values of 4.4 to 7.4 and thus outperforms the others in almost every instance. Those forecasts again can for example be used to regulate sources of pollutants such as, e.g., traffic.