Understanding and Eliminating Forecast Busts in Global Numerical Weather Prediction Models

This study investigates forecast failures in global numerical weather prediction (NWP) systems that are likely due to failures to accurately capture the interaction between mesoscale convective systems (MCSs) and the jet stream over North America. While the skill of NWP systems has advanced at a rate of roughly one day per decade, these systems still experience instances where predictive skill is well below anticipated levels or completely lacking. Such forecast failures are often termed “dropouts” or “busts” and are often defined as a failure to accurately predict the upper-tropospheric flow patterns. For example, Rodwell et al. (2013) defined a forecast bust on time-scales of ~6 days when the normally high anomaly correlation coefficient (> 0.8) at the 500 hPa level between the observed and predicted flow decreases below reliable levels (< 0.4) and the area average of the root mean square error correspondingly increases to above 60 m. These busts are relevant to sub-seasonal and seasonal prediction as busts are often associated with the failure to predict major changes in flow regime across the middle latitudes and the Arctic, such as a change in the sign of the North Atlantic Oscillation (Lillo and Parsons 2017) or failure to predict the onset of a Rex block over Europe (Rodwell et al. 2013). Decreasing the number of busts will result in significant benefits to society.

The cause of these forecast failures is still unresolved and a subject of current research. Rodwell et al. (2013) noted that the mean conditions for these forecast failures were associated with the presence of large persistent MCSs over the Great Plains of North America and proposed that a bust occurs when the outflow from the MCS amplified the downstream ridge of a Rossby wave packet and slowed the downstream propagation of the wave packet. Subsequently, Lillo and Parsons (2017) proposed that the cause of these busts were more diverse and due to the impact of MCSs, hurricanes, and winter cyclones interacting with the jet stream to initiate new Rossby wave packets, that are poorly represented in the model. In a more recent study, Rodwell et al. (2018) found that failures in the reliability of European Centre for Medium Range Weather (EMCWF) ensembles occur due to difficulties in capturing the ensemble spread in the interaction between MCSs and the jet stream.

This study seeks to understand how the interactions between observed MCSs and the jet stream lead to forecast failures. The period of interest for this study, June 2015, was during the Plains Elevated Convection at Night (PECAN) field campaign. PECAN is an ideal time frame to investigate busts since a total of seven forecast busts occurred during the project using the bust criteria and the ECMWF reanalysis modeling system utilized in Rodwell et al. (2013). Specifically, we explore the hypothesis of previous studies that the forecast model does not accurately represent the upper-level divergence associated with MCSs over North America. The meridional winds associated with this divergence could modify or trigger Rossby wave packets that would amplify across the Atlantic and “break” causing a large-scale forecast bust over Europe. This hypothesis will be investigated using observations from the nation’s operational (NEXRAD) radar network and comparing these fields against ECMWF forecasts and the ERA5 reanalysis. Following the approach of Homeyer et al. (2015), we utilize NEXRAD data to estimate the magnitude and height of the diverging outflow at the top of these convective storm systems as a function of time. This study examines the character of the divergence signal for both bust and non-bust days to determine if the nature of the convection influences whether a forecast bust will occur. One finding is that strong convection occurs early in the storm’s lifetime with divergence at heights of 17.5 to 20 km above sea level, well above the height of the jet stream. These high-level divergence signals are critical for characterizing climate change as the injection of water vapor from the convection into the stratosphere changes the earth’s radiation budget, Homeyer et al. (2017). Following the initial intense convection in the bust cases, the divergence at the level of the jet stream is long-lived, in excess of 15 hours, with pulses between strong divergence and weak/no divergence during the MCS lifecycle. In contrast, during null cases, the MCSs are shorter-lived so that after the initial intense divergence signal, a persistent divergence does not occur, and the pulsing is generally lacking.

Given that we found that the divergence associated with convective systems varies between bust and non-bust dates, our ongoing research seeks to understand why this observed difference in MCS structure and lifetime occurs. We are also examining how well these variations in MCS structure and longevity are represented in the ECWMF model to determine if the inability to represent these variations in convection are a cause of forecast busts. Since global NWP systems only partly resolve convective storms, this effort involves collaboration with the ECMWF to examine and ultimately propose changes to the convective parameterization scheme in the model. Such improvements in the parameterization of convection in numerical models has important implications on time-scales from short-range NWP to climate projections as Sherwood et al. (2010) noted that the treatment of convection in climate models is a major source of uncertainty in climate projections. In addition, our ongoing research is investigating that the error growth associated with these busts penetrations into the Arctic.