*James David Taylor1, Takumi Honda1, Arata Amemiya1, Yasumitsu Maejima1, Takemasa Miyoshi1
(1.RIKEN Research center for computational science)
Keywords:data assimilation, convection, numerical modelling
As we enter the era of post peta-scale computing, convective-scale NWP will be performed at increasingly higher model resolutions, using more sophisticated data assimilation (DA) schemes and advanced observational datasets. In this study we explore the implications of a regional scale numerical weather prediction system that implements a unique 30-second update for a 500-m grid, using observations from a phased array weather radar (PAWR). The impacts are examined for both analyses and rainfall forecasts. Sensitivity experiments performed with the horizontal localization scale parameter showed a rapid buildup in dynamical activity in the analyses from the start of cycling that promoted the initialization of spurious and overly active convection in forecasts, causing the model to rapidly lose forecast skill. These conditions were found to be the consequence of substantial discrepancies between the initial conditions and observations, that introduced large perturbations to the analyses during initial cycling, to generate an atmospheric state that was characterized by strong low-level winds and regions of high instability. These conditions remained fairly constant by the 30-second updating after a period of initial cycling, continuing to degrade forecast skill through overly intense buildup of convection. It was demonstrated that these conditions could be limited in the model by reducing the localization scale parameter to near model grid resolution, which acted to force initial conditions closer to the initial set of observations following the first update.