Towards high-resolution climate projections for Belgium: Combining convection permitting models with statistical downscaling techniques

Precipitation has a large impact on the society resulting in a high demand for precipitation projections for the future. Apart from average precipitation projections, extreme precipitation projections are also crucial as hazardous weather associated with extreme precipitation can cause great damages. However, performing projections for intense precipitation is difficult as the climate models currently used to perform these projections have rather poor performance for extremes. Due to the typical grid size of current climate models (i.e., up to 12 km), convective processes are represented by parametrizations resulting in deficiencies, notably underestimations of hourly precipitation intensity and misrepresentations of the diurnal cycle of precipitation. Models with a grid fine enough to partly represent dynamically convective processes (at least as fine as 4 km) show clear improvements. Until recently such models, also referred to as convection permitting scale (CPS) models, were used only in a weather forecasting framework due to their high computational cost. However, the increase in computational resources currently allows their integration over yearly or decadal time-scales. The question arises whether climate projections could benefit from CPS simulations and if the computational cost of such CPS- simulation based projections can be lowered. To assess the need for climate projections benefiting from CPS simulations, a first step is to ensure that CPS simulations show an added value compared to coarser simulations. This verification was performed with eleven-year COSMO-CLM simulations at different resolutions (25 km, 7 km and 2.8 km). It was found that the main biases occurring at non-CPS are corrected at CPS. The representation of the diurnal cycle of precipitation is improved, especially in the afternoon when convective activities reach a maximum. In addition, the representation of hourly precipitation is more realistic in the CPS simulation compared to the non-CPS ones. Finally, the spatial representation of both precipitation and temperature is more accurate in the finest resolution simulation. The need for CPS simulations in climate projections is also depending on the existence of significant differences between the CPS and non-CPS simulations. If CPS simulations show identical climate sensitivity than non-CPS simulations then a simple bias correction of the non-CPS simulation could provide similar performances and much lower computational cost than CPS models. To verify this, two additional simulations were performed at CPS using the EC-Earth model as lateral boundary conditions. No significant changes were found for precipitation’s spatial variability. However, the climate sensitivity for daily precipitation intensity quantiles greater than 20 mm/day was found to be 25% higher in the CPS simulations compared to the non-CPS ones. Such results show the need to include CPS simulations in climate projections to improve the range of uncertainty notably concerning the highest precipitation quantiles. Although climate projections could benefit from the use of CPS simulations, the current computational cost of CPS simulations does not allow the production of simulations ensembles at CPS. In this thesis, three approaches are developed to lower the computational cost of CPS simulations on the one hand and of climate projections on the other hand. These approaches consist of (1) the selection of a CPS model configuration that combines low computational cost and realistic representation of convective precipitation, (2) the quantification of the uncertainty related to the use of shorter integration periods compared to what is commonly done in current climate projections and (3) the development of a physically-based statistical framework that could be later used as basis for complimentary alternative computationally cheap statistical downscaling models (SDMs). (1) Three different options are investigated that are likely to reduce the computational costs required to perform CPS simulations, namely switching off the parametrization of graupel, changing the domain size or using different nesting strategies. It was found that, among these three options, only the use of an efficient nesting strategy has the potential for reducing the computational costs (up to 25% lower than the reference simulation in our study) without deteriorating the representation of convective precipitation. (2) Decreasing the length of the time-period for which the climate model is integrated can also help to decrease the computational cost of climate projections. However, a reduction of the integration period results in an increase of the uncertainty related to the climate variability. For precipitation averages in Westdorpe (The Netherlands) the uncertainty of 11% over a 30-year integration period increases to 18% for a reduced 11-year period. Depending on the expected difference between two simulations one could adjust the integration period of these simulations and lower their computational cost. (3) Another approach to lower the computational cost of climate projections based on CPS simulations is to use a combination of computationally cheap SDMs together with CPS models. The circulation type (CT) framework was developed as a possible basis for developing physically based SDMs. It was found that by using the CT only, a large part of the precipitation variability is explained for winter and spring. However for convectively active seasons and for short time-scales (e.g., days), additional predictors are needed. First investigations show that consistent relations between temperature and precipitation are found within individual CTs bringing confidence that the framework of CTs is a good basis to develop robust SDMs with statistical relationships that hold for different climates.status: publishe