Regional climate models are important tools in climate change impact studies due to their high horizontal resolution. On the other hand, regional simulations still include considerable uncertainties and can have substantial biases in comparison to observations. Thus, before the data can be used for deriving climate projections, these biases have to be identified and, to the extent possible, eliminated. There are two approaches to combine the information from observations and simulations: either to adjust observations with the simulated change (delta-change approach) or to correct the biases in the simulations relative to observations during a control period.
In this thesis, seven projection methods for daily precipitation were tested in a cross-validation framework. Model simulations taken from the ENSEMBLES data set were used to test the relative performance of these methods. In addition to traditional delta change method that scales only time mean precipitation, three algorithms which take daily variability into account were used. Two of these (Engen-Skaugen and power transformation algorithms) scale the standard deviation, while the most flexible one does the correction/adjustment percentilewise (analogy algorithm), so that changes in the shape of the distribution are also taken into account. The algorithms were applied both using delta change and bias correction approaches.
The performances of the projection methods depend on time, location and also the part of the distribution considered. Bias correction done with the analogy-algorithm worked well in a large part of the distribution, especially in north Europe. Due to smaller fraction of wet days and larger intermodel differences in the simulations, delta change methods performed relatively better in south Europe than in north Europe. On the other hand, bias correction with the power transformation algorithm has the best ability to adjust heavy precipitation, apparently due to the strong scaling it applies to the upper tail of the distribution. The results improved when the projections for the best performing methods were combined. The reason is the same as with multi model mean projections: errors in different projections tend to cancel each other out.
To assess the uncertainty due to intermethod differences, methods were applied directly to observations taken from the data set gathered by European Climate Assessment & Data. The results showed that most of the overall uncertainty (if only one emission scenario is used) comes from intermodel differences that are large especially for bias correction methods. Uncertainty related to intermethod differences is smallest in the middle parts of the distribution, but increases towards the tails of the distribution and tends to be largest in summer. Thus, the intermethod differences are non-negligible and should be taken into account when calculating daily precipitation projections.
