Bias correction of precipitation from convection-permitting models at the point scale: a case study in Switzerland

In comparison to regional climate models, convection-permitting models (CPMs) improve predictions of short-duration extreme rainfall by resolving convective processes. Their sub-daily rainfall output is thus valuable for impact models like those used for urban drainage systems. However, limitations remain, including short simulation periods and biases requiring correction to match station-scale resolution. To overcome these limitations, this study evaluates five quantile mapping (QM) approaches to bias correct and downscale sub-hourly COSMO-CLM simulations at 2.2 km resolution to the station scale using data from over 70 weather stations in Switzerland. To more effectively analyze the impact of QM on different rainfall features, stations were grouped based on location, altitude, and rainfall characteristics through an unsupervised clustering algorithm. Conventional QM was tested against several advanced methods, including using a moving window, spatial pooling, and extending the observational record. These techniques were validated using cross-validation and by evaluating several historical rainfall indices and the climate change signal. Results reveal that wet biases in raw CPM output remain and can exceed 90%. QM reduces biases in annual precipitation indices, but conventional QM often overcorrects, introducing dry biases in extreme quantiles. Additionally, all QM approaches can modify climate change signals, inflating or reversing trends in key indices. Combining spatial pooling with a moving window shows the most promise for reducing overall bias; however, challenges persist, such as data scarcity for extreme events, observation-simulation precision disparities, and short calibration periods. Future work should explore seasonality and utilize diverse CPM ensembles for more reliable precipitation projections.