Multivariate post-processing of sub-seasonal weather regime forecasts

Reliable forecasts of quasi-stationary, recurrent, and persistent large-scale circulation patterns – so-called weather regimes – are crucial for various sectors of society, including energy, health, and agriculture. Despite steady progress, probabilistic weather regime predictions still exhibit significant biases and are not reliable beyond 15 days of lead time. Thus, this study aims to advance their predictions through ensemble post-processing. Our approach is based on a year-round regime definition that distinguishes between four types of blocked regimes dominated by high-pressure situations in the North Atlantic-European region and three types of cyclonic regimes dominated by low-pressure situations. The manifestation of each regime can be expressed by a seven-dimensional weather regime index representing the projection of the 500-hPa geopotential height field onto the mean patterns of the seven weather regimes.
This index is calculated for ECMWF’s sub-seasonal reforecast ensemble data valid in the period 1999 to 2020 and verified against ERA5 reanalyses. To improve the accuracy and reliability of the multivariate probabilistic weather regime forecasts, we adjust the raw model outputs respective to their uncertainties and biases using a combination of Ensemble Model Output Statistics (EMOS) and Ensemble Copula Coupling (ECC). With EMOS, the year-round mean skill horizon (referring to the 0.1 level of the CRPSS compared to the climatological forecast) increases by 1.5 days compared to the current state-of-the-art weather regime forecast. We further replace the univariate EMOS method with a neural network-based distributional regression approach that provides greater flexibility in predictor intake.
Overall, our study reveals that statistical post-processing techniques are one way to improve weather regime forecasts, which can help plan and manage, reduce risks, and maximise societal benefits.