Probabilistic Precipitation Forecasts from Convection-Allowing Models and Ensembles

The new generation of convection permitting NWP models has added unprecedented detail and realism to precipitation forecasts, particularly for convective rain. However it has also become apparent that the spatial accuracy of the forecasts does not necessarily live up to the heightened expectations, especially for high impact events. The precipitation may look convincing but is often in the wrong place due to limited predictability at local scales. Hence, in order to make best use of these forecasts, they should really be interpreted and presented probabilistically.

Probabilistic forecasts can be produced either by post-processing the deterministic forecast, combining the information from recent forecasts (time-lagging), running an ensemble, or by a combination of these methods. One way of obtaining probabilities (or scenarios) from a deterministic forecast is neighborhood processing. A neighborhood (a square or circle of a particular size) is centred around each model pixel and the fraction of occurrences of an event (e.g. rain above 2mm/hr) in that neighborhood is interpreted as the probability that that event will happen at the central pixel. Thus it is assumed that anything which happens in the locality of a point is equally likely to happen at that point. This approach significantly improves forecast skill but raises the question how large should the neighborhood be? This question will be addressed.

In response to the local predictability issue, several forecast centres including the UK Met Office now run convection-permitting ensembles. The Met Office ensemble (MOGREPS-UK) currently runs 12 members every 6 hours out for 36 hours on a domain covering the UK with a grid spacing of 2.2km. This is a big step forward in capability, but for precipitation forecasts, 12 members are not sufficient to fully capture the local spatial uncertainty, resulting in noisy probabilistic forecasts with spurious unreliable detail. Smoother and more reliable probabilities can be obtained by the additional use of neighborhood processing, but this re-introduces the question of which neighborhood size is optimal.

A new method has been developed that makes use of the ensemble members to generate neighborhood sizes that can vary with forecast time and model grid square. This adaptive approach can generate small neighborhoods where ensemble members are in close spatial agreement and larger neighborhoods where there is greater spatial uncertainty and should therefore give improved probabilistic forecasts if the underlying ensemble is reasonable. This method has the advantage of not needing to be tuned, and doesn't impact the larger-scale ensemble spread.

The presentation will introduce the need for neighborhood processing of precipitation forecasts for both the deterministic Met Office 1.5km UK model and the MOGREPS-UK ensemble. Probabilistic skill will be compared for different sized neighborhoods for forecasts over an eight month period in 2013. The visual effect on probabilities by applying the adaptive neighborhood will be shown and the skill obtained by using this approach compared with that from fixed neighborhoods. The results provide insight into the spatial predictability of precipitation from both the deterministic and ensemble systems and how to produce improved probabilistic forecasts.