Exploring metrics of extreme daily precipitation in a large ensemble of regional climate model simulations

The ability of a large ensemble of fifteen state of the art regional climate models (RCMs) to simulate precipitation extremes is investigated. The 99th, 99.9th, and 99.99th percentiles of daily precipitation in the models are compared to those in the recently released E-OBS database for the winter, spring, summer and autumn season. The E-OBS database contains daily, area averaged precipitation amounts, covering Europe from 1950 to 2008 at 25 and 50 km resolution. It is found that the majority of the models overestimate the values of the precipitation extremes compared to E-OBS, on average by about 38%, but for some models even exceeding 50% in the European mean. To measure the model performance, a simple metric is proposed which averages a nonlinear function of the seasonal biases over the European area. The sensitivity of the metric to different assumptions in the construction and the quality of the observational data is explored. Generally low sensitivities of the metric are found to spatial and seasonal averaging. However, large sensitivities are found to potential biases in the observational atabase. An alternative metric that measures the patial pattern of the extremes (which is not sensitive to a potential constant offset in the observational data) is further explored. With this metric the ranking between the models changes substantially. However, the two models with the worst score in the standard metric also display the worst scores with this alternative metric. Finally, it is shown that the regional climate models display the largest biases compared to E-OBS in areas where he underlying station density used in E-OBS is low, thus suggesting that data quality is indeed an mportant issue.

Summarizing the results show that i) there is no metric that guarantees an objective and precise ranking or weighting of the models, ii) by exploring different metrics it nevertheless appears possible to indentify models that perform consistently worse than other models, and that iii) the observational data quality should be considered when designing and interpreting metrics.