Bayesian multi-model projections of climate: generalization and application to ENSEMBLES results

In a previous study, we developed a Bayesian methodology for combining multi-model climate change simulations into a single probabilistic projection which addresses changes in inter- annual variability beyond changes in mean temperature, and which explicitly considers time-depen- dent model biases. We tested 2 different but equally plausible bias assumptions referred to as 'con- stant bias' and 'constant relationship'. The former assumes that the biases in control and scenario periods are approximately constant, following the implicit assumption in most climate change stud- ies. The latter approach follows seasonal forecasting procedures by assuming an approximate linear relationship between observed and simulated seasonal temperatures. In the present study we gener- alized this approach by combining the 2 bias assumptions into a single probabilistic projection. In cases where the 2 assumptions yield conflicting results, our methodology implicates a broader prob- ability density function, thereby reflecting the increased level of uncertainty. We applied the new method to area-mean seasonal temperature distributions from global/regional climate model simula- tions of the ENSEMBLES project. Results are presented for changes in mean and variability between control (1961-1990) and scenario (2021-2050) periods. In comparison to the multi-model mean, the generalized Bayes method projected a considerably weaker warming during summer and autumn in much of continental Europe, a stronger winter warming in Scandinavia, France, eastern and central Europe, and a weaker warming in both summer and winter in the Mediterranean. These differences can be traced back to the models' difficulties in representing the natural interannual variability in these regions.