ECMWF has hosted a workshop on ‘Stochastic Parametrization’ which reviewed progress in the simulation of uncertainties in forecasting models and examined the scope for further improvements.
The event, which took place on 9 and 10 March, was organised by Prof Tim Palmer from Oxford University, and Dr Antje Weisheimer, who works for both ECMWF and Oxford University. It brought together about 40 participants, mainly from ECMWF, Oxford University and the UK Met Office.
Prof Palmer, one of the first three ECMWF Fellows appointed in 2014, said stochastic parametrization was becoming “a more and more important technique, not only for weather prediction, but also for climate prediction”.
“In order for ECMWF's forecasts to be useful for quantitative decision-making, it is essential that they come with reliable estimates of forecast uncertainty. The computational models used to make the forecasts are themselves only approximations of reality, and hence are a source of forecast uncertainty. Stochastic parametrization is an important technique to represent model uncertainty quantitatively in an ensemble forecast system,” he said.
An ensemble forecast system predicts a whole set of possible outcomes, taking into account uncertainties arising from our imperfect knowledge of the current state of the atmosphere and from limitations in the forecasting model.
Stochastic schemes are now used routinely in the atmosphere component of ECMWF’s ensemble forecasts and are being developed and tested in the land, ocean and sea-ice components. The schemes simulate model uncertainty by introducing stochastic perturbations into the equations of the forecasting model.
The term ‘stochastic’ means that we cannot precisely determine these perturbations and instead define them in statistical terms. In the ensemble forecast context, the perturbations also have spatial and temporal structures selected to simulate as accurately and reliably as possible the effect of model uncertainty on forecast quality.
The contours in the figure below show examples of perturbation patterns in one of the two schemes used in ECMWF ensembles. This scheme introduces stochastic perturbation at three different spatial and temporal scales.
Examples of perturbation patterns at three spatial scales used in ECMWF medium-range/monthly (ENS) and seasonal (S4) ensembles
Stochastic parametrization is an area where ECMWF did some pioneering work in 1998–99, after Prof Palmer started to work on the problem in 1996. Today, model uncertainty simulation using stochastic schemes is one of the main areas of collaboration between ECMWF, Prof Palmer and his group at Oxford University, and the UK Met Office.
This workshop provided an opportunity for scientists working in this area to present ongoing work and to discuss preliminary results and new ideas. It was the third in a series started two years ago to promote collaboration and the exchange of information. The first workshop was held at Oxford University in 2013, followed by a workshop at the Met Office last year.
Many speakers at this third workshop showed how aspects of model behaviour can be improved by using stochastic schemes, but they also indicated that the way models react depends on the models themselves. Thus, although it is clear that stochastic schemes improve probabilistic forecasts, it is difficult to draw general conclusions on which scheme performs best, and on their relative benefits. The next few years should provide more clarity on whether there is a 'best approach' that can be followed to further advance the simulation of uncertainties in forecasting models.