Domain decomposition in shallow lake modelling for operational forecasting of flooding

The Netherlands is a highly urbanized area. In addition to flooding from the sea due to storm surges and high water discharges from rivers, flooding from major lakes is also a threat. Since 2011 there is a new system in operational use (24 hours per day, 7 days per week), for the prediction of flooding at Lake IJssel, Lake Marken, and the lakes bordering them. This system, RWsOS Meren (Genseberger et al. [2013]) enables a real-time dynamic forecasting of wind driven waves, water flow, wave runup, and overtopping at dikes. At the moment the time horizon of forecasts with RWsOS Meren is two days ahead. To enlarge this time horizon, medium-range global weather forecasts from ECMWF (ECMWF) up to 15 days (two forecasts per day) and short-to-medium range forecasts of extreme and localised weather events from COSMO-LEPS (limited area ensemble prediction system) (COSMO) up to 5, 5 days (one forecast per day) will be used as input for RWsOS Meren. In RWsOS Meren, only the two shallow-water models of the lakes will be run with this input (and not the models for waves, wave runup, and overtopping). ECMWF and COSMO-LEPS use ensembles (51 and 16 ensemble members, respectively). Therefore, also the two shallow-water models will be run in ensemble mode. As a consequence, for these models 204 runs with a simulation period of 15 days and 32 runs with a simulation period of 5.5 days have to finish within a reasonable time on a daily basis. This asks for a balance between low computational times per ensemble member and the efficient use of the available hardware (and energy) resources. In this paper we investigate how to manage this on current hardware.