Advanced statistical post-processing of ensemble weather forecasts

* Motivation *
Current weather prediction relies on complex numerical models of atmospheric circulation based on the fundamental laws of hydrodynamics and thermodynamics. Forecasts are made by dynamical ensemble prediction systems; to account for the uncertainty in initial conditions an ensemble of initial states is propagated under the model. Despite impressive improvements in the forecast skill of numerical weather prediction in the past decades there are still limitations due to model error and problems in generating ensembles. Model error may be addressed using multi-model ensembles or by stochastic parametrization. Nevertheless, it is observed in the Met Office's practice that forecast ensembles are still biased both in location and dispersion (see also, e.g., Hamill and Colucci, Mon. Wea. Rev., 1997; Gneiting et al., Mon. Wea. Rev., 2005; Raftery et al., Mon. Wea. Rev., 2005). They tend to be underdispersive, leading to overconfident uncertainty estimates and an underestimation of extreme weather events. Systematic biases are significant in subgrid-scale weather phenomena such as UK temperature, precipitation or wind speed at particular locations and state-of-the-art systems occasionally miss extreme weather events within the ensemble distribution. The raw ensemble distribution can thus not be expected to convert directly into a predictive distribution for a variable of interest.

* Statistical post-processing *
This leads to the idea of combining dynamical and statistical information to improve prediction by statistical post-processing of the dynamical ensemble. Proposed methods range from simple model output statistics schemes known since the 1970s to more advanced approaches such as ensemble dressing (Roulston and Smith, Tellus A, 2003; Wang and Bishop, Q. J. Roy. Meteor. Soc., 2005), Bayesian model averaging (Raftery et al., Mon. Wea. Rev., 2005) and non-homogeneous Gaussian regression (Gneiting et al., Mon. Wea. Rev., 2005). A recent study (Hemri et al., Geophys. Res. Lett., 2014) based on historical ECMWF forecast data has shown that the forecast skill gain over the raw ensemble due to post-processing remains constant even if the skill of the raw ensemble itself increases due to better forecast models and better data assimilation procedures. This indicates that post-processing will add skill for the foreseeable future. Until now, research on statistical post-processing has focussed on the average case, there has been little mention of rare or extreme weather events which are of high socio-economic impact.

* Project strategy *
The project will tailor existing and develop new methods for statistical post-processing of forecast ensembles with a particular view on extreme weather events. We will develop the promising novel approach of state-dependent post-processing. The post-processing will be conditional on the large-scale circulation regime the forecast model is in. We will use the Met Office's existing catalogue of weather regimes for this purpose. We will use historical data from the Met Office's ensemble prediction system MOGREPS together with the corresponding verifications. We are interested in short- to medium-range weather forecasting where there is considerable variability but still some skill in the ensemble. The research will be conducted in close collaboration with the Met Office as CASE partner. The project has the potential to produce key academic publications as well as real improvements in operational prediction capacity for extreme weather events.

* Objectives *
The main objectives of the project are:
(i) to develop and explore novel methods for statistical post-processing of forecast ensembles for extreme events;
(ii) to improve probabilistic prediction of extreme UK temperature, surface pressure, precipitation and wind speed;
(iii) to help implement better techniques in the Met Office's operational post-processing suite in order to improve prediction of extreme UK weather events.