Probabilistic evaluation of decadal prediction skill regarding Northern Hemisphere winter storms

Abstract

Winter wind storms related to intense extra-tropical cyclones are meteorological extreme events, often with major impacts on economy and human life, especially for Europe and the mid-latitudes. Hence, skillful decadal predictions regarding the frequency of their occurrence would be of great socio-economic value. The present paper extends the study of Kruschke et al. (2014) in several aspects. First, this study is situated in a more impact oriented context by analyzing the frequency of potentially damaging wind storm events instead of targeting at cyclones as general meteorological features which was done by Kruschke et al. (2014). Second, this study incorporates more data sets by analyzing five decadal hindcast experiments – 41 annual (1961–2001) initializations integrated for ten years each – set up with different initialization strategies. However, all experiments are based on the Max-Planck-Institute Earth System Model in a low-resolution configuration (MPI-ESM-LR). Differing combinations of these five experiments allow for more robust estimates of predictive skill (due to considerably larger ensemble size) and systematic comparisons of the underlying initialization strategies. Third, the hindcast experiments are corrected for model bias and potential drifts over lead time by means of a novel parametric approach, accounting for non-stationary model drifts. We analyze whether skillful probabilistic three- category forecasts (enhanced, normal or decreased) can be provided regarding winter (ONDJFM) wind storm frequencies over the Northern Hemisphere (NH). Skill is assessed by using climatological probabilities and uninitialized transient simulations as reference forecasts. It is shown that forecasts of average winter wind storm frequencies for winters 2–5 and winters 2–9 are skillful over large parts of the NH. However, most of this skill is associated with external forcing from transient greenhouse gas and aerosol concentrations, already included in the uninitialized simulations. Only over East Asia and the Northwest Pacific, the Northwest Atlantic as well as the Eastern Mediterranean the initialized hindcasts perform significantly better than the uninitialized simulations. While no significant differences are evident between anomaly- and full-field-initialization, initializing the model's ocean component from GECCO2-ocean-reanalysis yields slightly better results than from ORA-S4, especially over the Northeast Pacific. Additionally, it is shown that the novel parametric drift-correction approach – estimating potential cubic drifts with parameters linearly changing in time – is more appropriate than the standard procedure – estimating constant model drifts via the lead-time-dependent bias – and, hence, yields higher skill estimates.