Exploiting the benefits of convective-scale ensemble forecasts

Convection-scale ensemble forecasts are now widely used in operational centres and the research community to provide an estimation of uncertainty in the location and the severity of high-impact weather. These forecasts are run in regional domains with grid spacings of ~2km using much coarser resolution global forecasts (~10-20km) to provide lateral boundary conditions; the individual ensemble members differ in their initial conditions, boundary conditions and model physics.

The project aim is to investigate the factors controlling the spread of convective-scale ensembles at multiple spatial scales and under different flow regimes, and so improve the skill of short-range weather forecasts.

While their production is now routine, the techniques for initialising the multiple forecasts are still not firmly established. The Met Office, in particular, is exploring new strategies for creating perturbed forecasts e.g. they have recently changed their convection-permitting ensemble forecasting system (MOGREPS-UK) from a six-hourly-cycling to an hourly-cycling ensemble with multiple lag times. Their aim is to both provide more timely forecasts by reducing the time between assimilation cycles and forecast availability, and to increase the spread of the ensemble by taking into account additional perturbations coming from successive data assimilation (DA) cycles.

Similar to convective-scale ensembles produced by most operational centres, MOGREPS-UK is under-spread e.g. the calculated probabilities of precipitation are too high when compared to the forecast error across many events. However, different types of perturbations may also lead to unwanted spread and convective-scale ensemble behaviour is likely to depend on the weather regime i.e. scattered convective showers will behave differently to organised convection within a weather front.

In this project recently-developed metrics for evaluating ensemble spread and spatially-skilful scales will be applied including the neighbourhood-based "Fractions Skill Score" (FSS), dispersion FSS (for ensemble dispersion), ensemble agreement scale and correspondence ratio. The consistency of spread interpreted from different metrics (including also standard objective verification metrics of spread/skill ratio and spectral analysis) and different output fields, and for different weather regimes, will be determined. Additionally, power spectrum analysis may be used to further inform the sensitivity of the ensemble spread to spatial scales.

This verification analysis could support the design of a well-spread ensemble based on new DA developments and perturbation types, as well as ensemble configuration. In the next few years MOGREPS-UK will be at the centre of several developments in the DA system, such as the change in the DA scheme for the global ensemble MOGREPS-G (from ETKF to En-En4DVar). On top of the perturbations associated with the large scales, new developments will also be taking place in convective-scale DA including a new DA system, hybrid-UKV. These developments in convective-scale ensemble design provide us with the opportunity to investigate the response of the ensemble spread and skill. The long-term goal is to optimally design convective-scale ensembles in terms of perturbation type, ensemble size and verification metrics with emphasis on the assessment of spread at small scales.