Detection of potential extreme precipitation events in convection-permitting climate simulations using statistical methods in global climate models

Extreme precipitation is one of the most dangerous meteorological phenomena, causing significant harm to people, structures, and crops, as well as resulting in economic losses. Due to their resolution and configuration, Convection-Permitting Models (CPMs) offer the potential to reproduce convective extreme precipitation. However, CPM simulations require long runtimes and generate large amounts of data, consuming significant computer storage. Therefore, it is necessary to establish a link between extreme events in CPMs and their counterparts in global climate models (GCMs) and regional climate models (RCMs). Identifying this link would allow selective use of CPMs only for days when significant precipitation is expected, reducing the need to run climate simulations over extended periods and focusing on specific days.

Three methods were tested for detecting potential extreme precipitation events in CPMs using GCM data: a fixed precipitation threshold, linear regression, and logistic regression. Convective precipitation and calculated instability indices (Lifted Index, K Index, CAPE, wind shear, moist convergence and wind convergence) were used in the selected GCM, EC-Earth. The methods were trained using the outputs of the CPM HCLIM38-AROME over the Scandinavian domain in the historical period (1986–2005). This domain was chosen for training because of the availability of longer time series, and it was further divided into three smaller subdomains for improved method evaluation. Additionally, RCM simulations (HCLIM38-ALADIN) were used to enhance the detection of extreme events. The RCM bridges the gap in grid spacing between the GCM and the CPM. For validation, standard statistical methods were used, such as hit rate, Matthews correlation coefficient, and ROC score.

Initial results using a convective precipitation threshold in the GCM show relatively successful forecasting of extreme precipitation in the CPM, with a hit rate of 0.42. Calculating instability indices and introducing linear and logistic regression further improve the results, increasing the hit rate to 0.51. When moving to smaller subdomains within the same period, the performance of all methods decreases somewhat, but linear and logistic regression remain effective in predicting days with extreme precipitation (hit rate varies from 0.35 to 0.43).

The next step was to introduce the intermediate RCM as an additional filter. With this modification, the extreme precipitation prediction results reach a hit rate of 0.65.

Finally, the methods are validated over the same domain for the same models in the future period (2081–2100), as well as over the pan-Alpine domain, which includes parts of the Mediterranean and Croatia, in the period 1996–2005. The results indicate that by using linear regression and the RCM as a filter, extreme precipitation events can largely be detected already in the GCM, allowing selective inclusion of the CPM, which leads to savings in computational resources and time.