Summary: | While the average skill of medium-range numerical weather prediction (NWP) has steadily improved over the last three decades, there is still considerable variance in day-to-day forecast performance. Much of this variance is contained within a long tail in the distribution that is skewed toward cases with very low skill, often referred to as forecast busts or dropouts. These forecast busts in global models are typically focused on sub-continental scales and can be associated with poorly-predicted high-impact weather events, motivating efforts to understand why these busts occur, how they could be anticipated, and how forecast systems could be improved to reduce their occurrence. This study is a systematic investigation of the variability in both upscale error growth and error propagation in global NWP. Our approach utilizes a framework for diagnosing error growth that begins with a prognostic equation for potential vorticity (PV) error in which non-linear terms have been mathematically eliminated. Following adiabatic flow, a wave equation is derived for the wind and PV error from which diagnostics for wave amplitude, wave activity flux (WAF), and Rossby wave source are defined. These diagnostics are then applied to ten years of deterministic ECMWF forecasts. Our results show that in the first 24 hours the largest rotational errors at the tropopause are over the central US, and to a lesser extent eastern Asia, during the spring and summer. These errors subsequently expand downstream within the respective waveguides. During the winter, initial error growth shifts to the eastern Pacific. The difference between good an bad medium range deterministic forecasts for Europe is associated with error growth over North Atlantic. To further investigate this issue, MPAS forecast runs are presented for cases during increased MCS activity over the central US during June 2015. This period coincided with the PECAN (Plains Elevated Convection at Night) field campaign and also included multiple forecast busts in the ECMWF model. ...
|