New approaches to postprocessing of multi‐model ensemble forecasts

Abstract

Ensemble weather forecasts often under‐represent uncertainty, leading to overconfidence in their predictions. Multi‐model forecasts combining several individual ensembles have been shown to display greater skill than single‐ensemble forecasts in predicting temperatures, but tend to retain some bias in their joint predictions. Established postprocessing techniques are able to correct bias and calibration issues in univariate forecasts, but are generally not designed to handle multivariate forecasts (of several variables or at several locations, say). We propose a flexible multivariate Bayesian postprocessing framework, based on a directed acyclic graph representing the relationships between the ensembles and the observed weather. The posterior forecast is inferred from available ensemble forecasts and an estimate of the shared discrepancy, obtained from a collection of past forecast–observation pairs. We also propose a novel approach to selecting an appropriate training set for estimation of the required correction, using synoptic‐scale analogues to obtain a regime‐dependent estimate of the adjustment. The proposed technique is applied to forecasts of surface temperature over the UK during the winter period from 2007 to 2013. Although the resulting parametric multivariate‐normal probabilistic forecasts are marginally less sharp than those of the leading competitor, they capture the spatial structure of the observations better than a correlation structure based on either the ensembles or climatology alone, and are robust to changes in the variables and spatial domain of the forecast, at a greatly reduced computational cost.

Recalibrating wind‐speed forecasts using regime‐dependent ensemble model output statistics

“…Regime-dependent methods with indicator weights can therefore also be interpreted as analogue-based post-processing approaches, whereby a training data set is constructed from forecast-observation pairs that are believed to exhibit behaviour similar to the current forecast (Junk et al, 2015). In this case, the assumption is that the forecast biases depend on the synoptic-scale behaviour of the atmosphere, which aligns with the motivation for using regime analogues in Barnes et al (2019).…”

“…The number of principal components chosen is p ≪ 1024; the leading three principal components are retained here, explaining 22.0% of variation in the hemispherical streamfunction. Barnes et al (2019) define weather regimes as the leading principal components of mean sea-level pressure fields over a relevant spatial domain. In the work presented here, the synoptic-scale atmospheric state is similarly projected on to the leading principal components, but regimes are then identified by performing an additional clustering step in this reduced space.…”

Remember the past: A comparison of time-adaptive training schemes for non-homogeneous regression

“…Alternative time-adaptive models are based on historical analogs or non-parametric approaches. For approaches employing analogs (Junk et al, 2015;Barnes et al, 2019), training sets are selected to consist of past forecast cases with atmospheric conditions similar to those on the day of interest. Such methods may lead to models that are able to account for the flowdependency of EPS errors (Pantillon et al, 2018;Rodwell et al, 2018).…”

Which precipitation forecasts to use? Deterministic versus coarser‐resolution ensemble NWP models

“…A natural idea is to combine these two forecasts appropriately to generate new forecasts that are better than either one across all locations. Many methods have been developed to merge available data information from different sources (Rodwell, 2006; Wang et al ., 2012; Barnes et al ., 2019; Xu et al ., 2019; Leutbecher and Ben Bouallègue, 2020). It will be useful to evaluate how to apply some of these techniques to combine these two types of NWP forecast to take advantage of their strengths.…”