Extreme winds and sea-surges in climate models

This thesis deals with the problem of how to estimate values of meteorological parameters that correspond to return periods that are considerably longer than the length of the observational data sets. The problem is approached by considering the output of weather- and climate models as pseudo-observations. These pseudo-observational records, which are one to two orders of magnitude longer than the observational records, open the possibility to reduce the large statistical uncertainty in the 104-year estimate from observations, as well as to examine the assumption that all extremes (up to 104-year return periods) are part from the same population. In Chapter 1 we quantify the statistical uncertainty in the 104-year surge level if estimated from hundred-year records (as the observational records are). This is done by dividing the 5336-year long outputs of the climate model ECBilt-Clio into subsets of hundred year. This chapter shows that annual maxima of hundred-year surge records can generally, within the uncertainty, be described by a Gumbel distribution (a commonly applied distribution for annual hydrological extremes (see e.g., Katz et al. 2002)). However, the total 5336-year record of the control run (1960-1990) can clearly not be described by a single Gumbel distribution, but requires a GEV distribution instead. This implies that uncertainty ranges calculated from Gumbel distributions will produce numbers that are misleadingly low (see also Coles et al. 2003). The uncertainty in the estimate of the shape parameter of the GEV distribution on basis of hundred-year records results in a large uncertainty (+- 4 m) in the 104-year surge estimate. For the grid point representing the North Sea, the greenhouse run (2050-2080) of ECBilt-Clio reveals a kink in the distribution of the annual maxima if displayed on a Gumbel plot. The extremes with a lower probability than once in 250 years seem to originate from another distribution than the less extreme events. It is hypothesized that the super-extremes originate from a second population in the extreme wind speed distribution. Chapter 2 deals with the optimal method of determining 104-year surge estimates by statistical means. Next to the GEV analysis (that usually only considers the annual maxima) the so-called Peak Over Threshold (POT) method exists, which considers all independent events above a certain threshold. These methods are evaluated with the ECBilt-Clio record of simulated surges in Delfzijl by first estimating the 100-year surge level and its uncertainty for all 116-year subsets, and then checking if these uncertainty intervals contain the correct realization, as determined directly from the total (7540-year) set. We found that in our experimental setting, the POTmethod systematically underestimates the uncertainty in the 102-year surge level, while application of the GEV distribution results in a unbiased estimate, making the last approach most appropriate for determining safety levels. Chapter 3 focuses on the superstorms detected in chapter 1. A statistical criterion is developed to determine whether all annual extremes can be described by a single GEV distribution or not. We found that for specific geographical locations in ECBilt-Clio, the extreme winds can not be described with a single GEV distribution, but requires a Generalized Two-Component Extreme Value (GTCEV) distribution. The meteorology resulting in the second component of the GTCEV distribution has the following characteristics: the extreme winds are related to situations in which two vortices merge into a single one. In addition, the cyclones are embedded in a strong jet stream, and extreme precipitation accompanies the development of the cyclone. It is found that the area for which a second population is detected shifts due to the greenhouse effect from the North-Atlantic ocean to the European continent. This explains that in chapter 1 the superstormsâ are only detected in the greenhouse run. In Chapter 4 we explore the suitability of the European Centre for Medium- Range Weather Forecasts (ECMWF) seasonal forecast archive for extreme value analysis of surges. The combined seasonal forecasts of the ECMWF cumulate to 1600 years. The high resolution in time and space and the more complete physics (even compared with state-of-the-art climate models) make these data highly appropriate to be analyzed with extreme value statistics. The results for the surge in Hoek van Holland shows good statistical agreement with the observed extremes. The long model record reduces the statistical uncertainty in the 104-year estimate with no less than a factor four. We demonstrate in Chapter 5 that the archived ECMWF seasonal forecasts can also be used for extreme value estimates of other variables than wind and surge only. Four examples are presented, i.e., the Rhine discharge at Lobith, the sluicing of Lake IJssel water into the sea, the closure-frequency of the Maeslant-barrier, and the (wave and sea level dependent) load on the Pettemer sea wall. The examples illustrate that the -still expanding- ECMWF data set offers unforeseen possibilities in modeling (hydrological) extremes. Especially, the simultaneous modeling of multiple extremes opens new perspectives. Preliminary results obtained with the so-called Challenge data are presented in Chapter 6. De superstorm that we analyzed in the Challenge data has similar characteristics as the events in the ECBilt-Clio model. This result supports the idea that the earlier detected superstorms are not a model-artifact, but rather seems to be indeed a phenomenon belonging to the real world.