Probabilistic Forecasting of Heavy Rainfall using Evolutionary Program Ensembles

Analyses indicate that heavy rainfall events have increased in recent years, and are expected to increase further with ongoing climate change. Such events are particularly damaging in built environments, whose infrastructure is not engineered for the increased frequency. From a warning and decision-support perspective, better forecasts of heavy rainfall are needed. Recent research has shown that a method known as evolutionary programming (EP) can produce large member ensemble forecasts, and that these forecasts provide greater probabilistic skill, particularly at the extremes, than traditional numerical weather prediction (NWP) ensembles. Further research has shown that this skill advantage persists out to longer ranges, where the forecast signal is presumably weaker. The EP approach, however, has not been applied to the difficult problem of precipitation forecasting or the further challenge of forecasting extreme precipitation events. This study presents a first approach to developing such forecasts.

Most heavy rainfall events in the midwestern U.S. develop from training but otherwise unremarkable convection. Past studies have established that training storms are often found on the cold side of a generally west-to-east quasi-stationary, frontal boundary and are typically characterized as nocturnal, elevated convection associated with a southerly low-level jet. The deep layer shear associated with these training systems is roughly parallel to the boundary. This indicates the importance of larger-scale atmospheric organization and suggests an opportunity to better quantify forecast uncertainty and to improve storm prediction. For example, the position of the training line will be sensitive to synoptic observational error, since relatively small spatial dislocations can determine whether or not a particular area will be affected and if so, how severely. The position error, in turn, depends on the dynamics of the stationary (usually, frontal) boundary and the low-level jet, each of which are likely to be highly sensitive to small errors in forecasts of the lower tropospheric wind field. The strength of the outflow boundaries from the convective storms themselves, which also govern storm evolution and movement, is a function of the energy available for storm development (convective available potential energy or CAPE). This can be assessed using available regional measures.

We use the unified precipitation dataset, with gridded precipitation at 0.25 x 0.25 degrees, for verification and apply the EP ensemble approach for the region from 40-45N and 88-92W (an area encompassing the region from central Illinois to central Wisconsin), for next day forecasts. Daily accumulated precipitation (P) for forecasts and observations are binned into five amount categories: P < 1.00 inches, 1.00 ≤ P < 1.50 inches, 1.50 ≤ P < 2.00 inches, 2.00 ≤ P < 3.00 inches, P ≥ 3.00 inches. Since a forecast of no measurable rain is relatively trivial, the training data are first screened to remove all observed days in which no measurable precipitation occurred. Input data to the EP include (1) maximum value in the time period of the regional average CAPE (obtained from RUC hourly analyses); (2) the time of the maximum CAPE; (3) regional average 0-6 km vertical wind shear (RUC) at the time of the maximum CAPE in (1); (4) regional average 700 hPa (mid-level) vector wind (RUC) at the time of the maximum CAPE in (1); (5) regional average 700 hPa (mid-level) vertical velocity (RUC) averaged over the time period; (6) regional average precipitable water obtained from RUC at the time of the maximum CAPE in (1); (7) the Model Output Statistics (MOS) QPF at a network of 16 regional NWS sites over each time period; and (8) the MOS QPF from the prior 1200 UTC forecast at the network of 16 regional NWS sites validating at the same time as the 0000 UTC MOS values. Additionally, measures of upstream, prior-day precipitation activity will be incorporated through rainfall measurements at specific sites. Preliminary results from this analysis will be presented.