Section: New Results
A strong limitation of chemistry-transport models lies in the high uncertainties in the physical parameterizations and the input data (meteorological data, emissions, ...). Uncertainties in input fields often range between 20% and 50%. Many physical parameterizations are available to estimate the same fields, which is usually the main source of uncertainties in output concentrations. In this context, a single forecast has little meaning: modern forecasting systems should plan to include ensemble forecasts, that is, simulations from an ensemble of models.
The modeling system Polyphemus has been designed in order to handle ensembles: it hosts several chemistry-transport models and includes many options with regard to the physical formulation of the models. An ensemble of 48 models has been studied to apply ensemble methods to photochemical forecasts (ozone). Methods have been proposed to build a forecast with linear combinations of the models outputs. Significant improvements in forecasts are obtained when the combination weights are optimal in least-square sense over the past (moving learning period).
On top of all models and of several model combinations, machine learning algorithms (sequential aggregation) have proven to improve performances. Unlike other methods, they come with theoretical bounds and are well known for their robustness. This is joint work with Gilles Stoltz (ENS Paris, DMA).
Ensemble forecasts (without combinations) have been tested daily (and for months) for the operational forecasting platform Prév'air (http://www.prevair.org/ ).