Calibrating ensemble reliability whilst preserving spatial structure

Abstract

A B S T R A C T Ensemble forecasts aim to improve decision-making by predicting a set of possible outcomes. Ideally, these would provide probabilities which are both sharp and reliable. In practice, the models, data assimilation and ensemble perturbation systems are all imperfect, leading to deficiencies in the predicted probabilities. This paper presents an ensemble post-processing scheme which directly targets local reliability, calibrating both climatology and ensemble dispersion in one coherent operation. It makes minimal assumptions about the underlying statistical distributions, aiming to extract as much information as possible from the original dynamic forecasts and support statistically awkward variables such as precipitation. The output is a set of ensemble members preserving the spatial, temporal and inter-variable structure from the raw forecasts, which should be beneficial to downstream applications such as hydrological models. The calibration is tested on three leading 15-d ensemble systems, and their aggregation into a simple multimodel ensemble. Results are presented for 12 h, 18 scale over Europe for a range of surface variables, including precipitation. The scheme is very effective at removing unreliability from the raw forecasts, whilst generally preserving or improving statistical resolution. In most cases, these benefits extend to the rarest events at each location within the 2-yr verification period. The reliability and resolution are generally equivalent or superior to those achieved using a Local Quantile-Quantile Transform, an established calibration method which generalises bias correction. The value of preserving spatial structure is demonstrated by the fact that 3)3 averages derived from grid-scale precipitation calibration perform almost as well as direct calibration at 3 )3 scale, and much better than a similar test neglecting the spatial relationships. Some remaining issues are discussed regarding the finite size of the output ensemble, variables such as sea-level pressure which are very reliable to start with, and the best way to handle derived variables such as dewpoint depression.

Combining predictive distributions for the statistical post-processing of ensemble forecasts

“…An entirely different approach to post-processing that completely circumvents the problem of choosing suitable parametric forecast distributions is the use of non-parametric methods, see for example Hamill and Whitaker (2006); Flowerdew (2014), and Taillardat et al (2016). However, these approaches suffer from the limitation that the support of the forecast distribution is restricted to the range of observed values in the training sets.…”

Generation of Scenarios from Calibrated Ensemble Forecasts with a Dual-Ensemble Copula-Coupling Approach

“…Non-representative correlation structures in the raw ensemble are magnified after calibration leading to unrealistic forecast variability. As a consequence, ECC can deteriorate the ensemble information content when applied to ensembles with relatively poor reliability as suggested, for example, by verification results in Flowerdew (2014).…”

Multivariate ensemble Model Output Statistics using empirical copulas

“…Examples of such multivariate settings include simultaneous forecasts of precipitation at networks of nearby locations (e.g. Bárdossy and Pegram, ; Flowerdew, ), forecasts for a temporally autocorrelated predictand, such as windspeed, at a sequence of lead times (e.g. Pinson, ) or (as will be used in this article) joint probability distributions for temperature and dewpoint to be used for computing predictive distributions for a ‘heat index’ (Steadman, ).…”