Ensemble forecasting with machine learning algorithms for ozone, nitrogen dioxide and PM sub(10) on the Prev'Air platform

This paper presents the application of an ensemble forecasting approach to the Prev'Air operational platform. This platform aims at forecasting maps, on a daily basis, for ozone, nitrogen dioxide and particulate matter. It relies on several air quality models which differ by their physical parameterizations, their input data and numerical strategies, so that one model may perform better with respect to observations for a given pollutant, at a given time and location. We apply sequential aggregation methods to this ensemble of models, which has already proved good potential in previous research papers. Compared to these studies, the novelties of this paper are the variety of models, the real operational context, which requires robustness assessment, and the application to several pollutants. In this paper, we first introduce the ensemble forecasting methods and the operational platform Prev'Air along with its models. Then, the sequential aggregation performance and robustness are assessed using two different data sets. The results with the discounted ridge regression method show that the errors of the forecasts are respectively reduced by at least 29%, 35% and 19% for hourly, daily and peak O sub(3) concentrations, by 19%, 26% and 20% for hourly, daily and peak NO sub(2) concentrations, and finally by 17%, 19% and 11% for hourly, daily and peak PM sub(10) concentrations. At last, we give a first insight of the ensemble ability to forecast threshold exceedances.