Comparing quantile regression methods for probabilistic forecasting of NO(2) pollution levels

High concentration episodes for NO2 are increasingly dealt with by authorities through traffic restrictions which are activated when air quality deteriorates beyond certain thresholds. Foreseeing the probability that pollutant concentrations reach those thresholds becomes thus a necessity. Probabilistic forecasting, as oposed to point-forecasting, is a family of techniques that allow for the prediction of the expected distribution function instead of a single future value. In the case of NO(2), it allows for the calculation of future chances of exceeding thresholds and to detect pollution peaks. However, there is a lack of comparative studies for probabilistic models in the field of air pollution. In this work, we thoroughly compared 10 state of the art quantile regression models, using them to predict the distribution of NO(2) concentrations in a urban location for a set of forecasting horizons (up to 60 hours into the future). Instead of using directly the quantiles, we derived from them the parameters of a predicted distribution, rendering this method semi-parametric. Amongst the models tested, quantile gradient boosted trees show the best performance, yielding the best results for both expected point value and full distribution. However, we found the simpler quantile k-nearest neighbors combined with a linear regression provided similar results with much lower training time and complexity.