A winter storm was forecast to impact much of the United States from the southern Plains to the Mid-Atlantic region from 2 to 3 March 2014. Forecasts from the GFS, EC, GEFS, and SREF converged on a potentially significant winter storm. Similar to the “Post-Groundhog Day Storm” and the March 2009 “Megastorm” (Stuart et al. 2013), despite a convergence of solutions of several models and ensemble forecast systems, the event proved to be difficult to forecast. The convergence of solutions and perhaps several relatively successful previous forecasts led to high confidence in the storm of 2-3 March 2014.
Dissimilar to the “Post-Groundhog Day Storm” and “Megastorm” (Stuart et al. 2013), the pattern of 2-3 March was generally well predicted, and a significant precipitation system developed farther south than all but short range forecasts. The frontal zone and the larger scale precipitation shield were poorly forecast and thus the snow potential on the northern edge of the frontal zone was never realized. The slow evolution of forecasts (Fig.1) closer to reality (Fig. 2) are illustrated using 6 European Center (EC) forecasts of total quantitative precipitation valid during the period of observed higher precipitation rates in the eastern United States (Fig.1). These forecasts, spanning 5 days, show the steady trend toward the axis of higher QPF shifting to the south. This trend will be shown to have impacted the National Centers for Environmental Prediction Centers (NCEP) models and ensemble forecast systems (EFS).
The EC model was shown here as it is often the model of choice by “social mediaforecasters” due to its perceived skill over other models and it is often the model of choice providing long lead-time forecasts on social media. Once these “forecasts” are released as described in Social Physics (Pentland 2014), the idea flow permeates a wide range of networked users and can have significant impact on decision makers and forecasters alike. These forecasts rarely account for uncertainty. When the forecasts are relatively accurate and consistent the collective intelligence may help get valuable information out. But when the forecasts are wrought with uncertainty, the collective intelligence may have a negative impact on the forecast process.
Accounting for uncertainty in forecasts is a complex issue. Ensemble Forecasts Systems (EFS: Sivillo et al. 1997), Lagged Averaged Forecast (LAF: Kalnay and Hoffman 1983; Dalcher et al. 1988), and a poorman’s ensemble (PME: Arribas et al. 2005; Bowler et al. 2008) are all viable methods to ascertain uncertainty in the forecast. Figure 1 is a simplistic example of a LAF, though it shows individual forecasts varying in time and the uncertain of these forecasts can only be visually inspected and estimated. This visual inspection method is often referred to as d(Prog/dt) making it natural to gravitate toward the trend. Hamill (2003) demonstrated that the trend visualized from d(Prog/dt) can and often does reverse, making it an unreliable forecast method. The spread and the probabilities derived from the three primary ensemble methodologies are the most useful means to ascertain the uncertainty information.
This paper will examine the winter storm of 2-3 March 2014. The overall pattern is presented, with a focus on where the precipitation fell and the pattern in the context of standardized anomalies. Forecasts presented focus on the uncertainty associated with forecasts of this storm. The focus is on the EC, NCEP GEFS and SREF forecast systems. The term forecast systems will be used in this paper to denote models and EFS-- all of which had difficulty with the forecast evolution of this event.
Methods and Data
The large scale pattern was reconstructed using 00-hour forecasts from the NCEP Global Forecast System (GFS). Standardized anomalies were computed as in Hart and Grumm (2001) using the re-analysis (Climate Forecast System; CFS) climate (R-Climate). The climatology spans a 30 year period. All data were displayed using GrADS (Doty and Kinter 1995).
Ensemble forecasts were derived from the NCEP Global Ensemble Forecast System (GEFS) and the Short Range Ensemble Forecast System (SREF). The surface and 500 hPa patterns were used to show how the general forecasts of a significant storm were predicted at longer ranges.
Snowfall data was retrieved from the National Snow site in text format, decoded in Python, and plotted using GrADS. The QPE data was retrieved from the Stage-IV 6-hour data. These data too were plotted using GrADS.
The GEFS mean was used to illustrate differences in the intensity of the 500 hPa heights between different GEFS cycles. The pattern was defined as forecast minus observed and, when using two forecasts, most recent minus older forecast cycle. Thus positive (negative) values imply a stronger (weaker) analysis or more recent forecast relative to the older forecasts.
The European Center (EC) model was retrieved from the EC TIGGE site for post analysis. It should be noted EC forecasts for snow were readily available on social media days before the event. LAFs were produced from the EC using GrADS to compute the mean and spread for all forecasts valid at the same time. The focus is on the run-to-run variability which helps capture the uncertainty in the forecasts.
. The 500 hPa pattern over the United States (Fig. 3) showed a deep trough over the eastern United Sates from 26-28 February (Figs. 3a-c) and a wave moving into California on 1 March (Fig. 3d). A ridge ahead of this trough produced above normal heights in the plains and southern United States as the wave moved across the country (Fig. 3d-f). At the surface (Fig. 4), two anticyclones dominated the pattern with a transient anticyclone moving across the United States from 26-28 February and a second massive anticyclone with +2 to +3 pressure anomalies moving out of northwestern Canada and into the United States (Fig. 4 c-f). This system was associated with a significant late season arctic air mass (Fig. 5) with pockets of sub -30C air at 850 hPa and -3 to -4 temperature anomalies (Fig. 6).
The anticyclone and the associated arctic air (Fig. 7) were critical components of both the forecast and the evolution of the precipitation shield. The shallow arctic air pushed deeper into the southern Plains than initially forecast, and it will be shown that the forecast systems all had difficulty with the strength of the arctic high and low-level cold air as well as problems with the Pacific wave.
Strong easterly flow, often associated with heavy rainfall and heavy snowfall (Stuart and Grumm 2008), did not develop with the evolving system until after 0000 UTC 3 March (Fig. 8) though a nearly textbook deformation zone was evident in the 850 hPa winds over Missouri at 1800 UTC 2 March 2014 (Fig. 8b). The strong southwesterly flow was present in the warm air as the system moved over the Mid-Mississippi Valley (MMV). Modest u-wind anomalies of -1 to -3 were present on the cold side of the frontal zone as the system moved across the MMV and Mid-Atlantic region.
A comparison of the winds to the QPE (Fig. 2), the heaviest precipitation fell on the warm side of the boundary over the MMV with areas of over 64 mm of QPE over northwestern Tennessee and Kentucky. In the easterly flow, where snow and sleet were observed, QPE amounts were generally under 16 to 20 mm. Deeper into the arctic air mass, QPE amounts were in the 2-8 mm range.
The precipitation (Fig. 2) during this event was closely linked to the frontal boundary. Due to the surge of arctic air, there were several interesting forecast aspects to the event, beyond the scope of this paper, to include the sleet which affected the southern Plains with elevated supercells producing hail in portions of Oklahoma1 during a sleet storm.
Forecast Systems Predictions
The primary forecast systems addressed here include the EC,NCEP GFS, NCEP GEFS and NCEP SREF. All EC forecast cycles form 0000 UTC 25 through 1200 UTC 1 March 2014 were retrieved but only 6 forecasts were shown to illustrate the evolution of the forecasts. The GFS and GEFS were used to show longer range forecasts from an NCEP perspective and the SREF was used to show shorter-range forecast issues. Beyond about 36 hours, the SREF performed poorly during this event as it had the arctic frontal boundary farther north than other forecast systems. These warmer solutions favored more freezing rain and sleet than was observed in locations such as Louisville, KY.
European Center forecasts 16 km (EC)
The QPF from 6 EC runs spanning the 5-days leading up to the event were shown in Figure 1. These EC forecasts, focused over the eastern United States showed the intial axis of 25 mm or more QPF to be well north of latter forecasts and the EC forecasts began to rapidly trend toward the observed QPF (Fig. 2) from forecasts initialized on and after 0000 UTC 1 March 2014. These robust QPFs in the cold air lead to optimistic forecasts of heavy snow in both social media and weather media space (not shown).
The corresponding EC forecasts of mean sea-level pressure (Fig. 9) showed the evolution of the anticyclone overtime and the surface cyclone. The anticyclone was stronger and pushed farther south in the shorter range forecasts and the surface cyclone was weaker, oriented more east-west and considerably farther south in shorter range forecasts produced by the EC.
The LAF for the EC mean sea-level pressure forecasts valid at 0000 and 1200 UTC 3 March (Fig. 10) shows the uncertainty on the cold side of the cyclone track and along the leading edge of the anticyclone as it pushed into the southern plains. The spread about the time averaged forecasts was 3 to 8 hPa over the Ohio Valley and portions of the southern Plains. Large spread in the LAF were indicated in the QPF, 500 hPa heights and 850 hPa temperatures (Fig. 11) over the region suggesting significant uncertainty in the frontal boundary where most of the QPF and wintery precipitation was forecast to occur.
Global Ensemble Forecast System 55 km (GEFS
The GEFS QPFs (Fig. 12) show the uncertainty associated with the probability of 25mm or more QPF from 6 GEFS forecasts. Similar to the EC deterministic forecasts, the GEFS showed the sharp trend to move the QPF shield to the south and east with 000 UTC 1 March 2014 being a critical transition time. The ensemble means QPF and each member 25mm contour (Fig. 13) show the same trend in the QPFs.
These data show the shift of the axis of higher QPF southward over time as did the EC forecasts. The mean seal-level pressure, temperatures and height fields (Fig. 14) all showed the large spread and high uncertainty in the frontal zone. The 500 hPa heights showed higher uncertainty with the Pacific wave before it was onshore and large spread about the mean with the pattern over Canada which diminished as the forecast horizon decreased with little spread apparent on and after 0000 UTC 1 March 2014 (Fig. 14e). The mesoscale frontal feature, where large gradients are typically present, was critical to correctly predict to get the QPF in the correct location.
Short-Range Ensemble Forecast System (SREF)
The SREF had larger differences than either the EC or GEFS though it had a similar overall error in its QPF fields (Fig. 15) Though not examined here, the SREF had significant precipitation type issues over the Ohio Valley into Pennsylvania which were not as strongly observed in the other forecast systems.
The SREF 15mm QPF probabilities from 0300 UTC 28 through 2100 UTC 1 March (Fig. 15) show the same general south and east trend in the QPF field over time. The 15mm contour was selected here as using a 10:1 snow to water ratio could be of value for forecast 6 or more inches of snowfall. These data show a rapid drop off in the QPF over 15mm over Pennsylvania with time an heavy snow as a low probability outcome much north of the Maryland border from forecasts issued on and after 2100 UTC 1 March 2014.
The SREF mean QPF and each member 25mm contour (Fig. 16) showed the larger trend observed in both the EC and GEFS. It is interesting to note that all 3 forecast systems showed the same general trend in the placement of the QPF shield and the timing of the changes were remarkably similar.
The SREF mean sea-level pressure forecasts and spread about the mean (Fig. 17) showed the same general pattern as the GEFS (not shown) and the EC based LAF (Fig 10). These data show a rapid decrease in the spread over time but a more significant decrease on and after 0300 UTC. The 0900 UTC time step was skipped though it showed the progression of decreased spread over time.
A strong anticyclone with attendant arctic air and a short-wave moving out of the Pacific ocean produced a widespread precipitation event from southern California to the East Coast. The impacts of the wave as it moved over California were not discussed here, with the focus being on the forecasts of the winter weather event across the United States east of the Rocky Mountains. This event was poorly predicted by the available forecast systems. The EC, GEFS, and SREF forecasts were presented here. The EC was used to produce a LAF showing the uncertainty issues associated with this event.
All three systems presented and all five systems examined (NAM and GFS not shown) had similar forecast issues and showed similar trends. It has been demonstrated (Hamill 2003) that the trend in normally not your friend as trends can and often do change. But in this event the trends were generally consistent and in general, the short term forecasts were relatively more accurate then longer range forecasts.
This case study reveals a number of science issues the must be addressed before more accurate forecasts of these types of events are possible.
Ensemble systems showed large spread indicating low confidence in the forecast
Forecast continually changed with each forecast cycle.
As the spread decreased forecasts became more accurate though not known apriori
LAF techniques helped show the uncertainty issues.
Generally poor forecasts of snow and ice were provided to decision makers
what tools could have helped assign low confidence to these forecasts
how to convey confidence and uncertainty to users
how do forecasters gain confidence indices of forecasts?
The Pennsylvania State University Department of Meteorology for data access and discussions related to this storm. Edited by Elyse Colbert.
Arribas,A, K.B Robertson, and K.R.Mylne: 2005 : Test of a poor man’s ensemble prediction system for short-range probability forecasting. Mon. Wea. Rev,133,1825-1839.
Associated Press 2014: Ice storm causes another traffic jam in the south. (and similar stories). Associated Press
Baker, B.W.1960: The 1960 ice storm in northern Alabama. Weatherwise, 13,196-200.
Bowler, N.E, A. Arribas,A, and K.R.Mylne: 2008 : The benefits of multianalysis and poor man’s ensembles. Mon. Wea. Rev,136,4113-4128.
Dalcher, A., and E. Kalnay, R.N. Hoffman, 1988: Medium Range Lagged Average Forecasts. Mon. Wea. Rev., 116, 402-416.
Doty, B. E., and J. L. Kinter III, 1995: Geophysical data and visualization using GrADS. Visualization Techniques Space and Atmospheric Sciences, E. P. Szuszczewicz and Bredekamp, Eds., NASA, 209–219.
Forbes G. S., R. A. Anthes, and D. W. Thompson, 1987: Synoptic and mesoscale aspects of an Appalachian ice storm associated with cold-air damming. Mon. Wea. Rev., 115, 564–591.
Gaza, B. and L. F. Bosart, 1990: Trough merger characteristics over North America. Wea. Forecasting, 5:314–331. [Abstract]
Grumm, R.H. and R. Hart. 2001: Standardized Anomalies Applied to Significant Cold Season Weather Events: Preliminary Findings. Wea. and Fore., 16,736–754.
Gyakum J. R., and P. J. Roebber, 2001: The 1998 ice storm—Analysis of a planetary-scale event. Mon. Wea. Rev., 129, 2983–2997. Find this article online
Hakim, G. J., L. F. Bosart, and D. Keyser, 1995: The Ohio Valley wave-merger cyclogenesis event of 25–26 January 1978. Part I: Multiscale case study. Mon. Wea. Rev, 123:2663–2692
Hamill, T.M, 2003: Evaluating Forecaster’s Rule of Thumb: A study of d(prog)/dt. Wea. Forecasting,18,933-937.
Harlin, B.W. 1952: The great southern glaze storm of 1951. Weatherwise,5,10-13.
Hart, R. E., and R. H. Grumm, 2001: Using normalized climatological anomalies to rank synoptic scale events objectively. Mon. Wea. Rev., 129, 2426–2442.
Hoffman,R.N. and E. Kalnay, 1983: Lagged Average Forecasting. Tellus, 35A,100-118.
Kocin, P. J., and L. W. Uccellini, 2004: Northeast Snowstorms, Volume I: Overview. Meteor. Monogr., Vol. 32, No. 54, Amer. Meteor. Soc., 1-296.
Kocin, P. J., and L. W. Uccellini, 1990: Snowstorms along the northeastern Coast of the United States: 1955 to 1985. Meteor. Monogr., No. 44, Amer. Meteor. Soc., 280p.
Irland L. C., 2000: Ice storm 1998 and the forests of the Northeast. J. For., 96, 32–40. Find this article online
Pentland, Alex: 2014: Social Physics: How Good Ideas Spread—The Lessons from a New Science. 320p. Penguin Press.
Rauber R. M., M. K. Ramamurthy, and A. Tokay, 1994: Synoptic and mesoscale structure of a severe freezing rain event: The St. Valentine's Day ice storm. Wea. Forecasting, 9, 183–208. Find this article online
Richwien, B. A., 1980: The damming effect of the southern Appalachians. Natl. Wea. Dig., 5(1), 2–12.
Robbins, C.C and J.V. Cortinas 2002: Local and synoptic environments associated with freezing rain in the contiguous United States. Wea. Forecasting,17,47-65.
Root, B., P. Knight, G.S. Young, S. Greybush, R.H. Grumm, R. Holmes, and J. Ross, 2007: A fingerprinting technique for major weather events. Journal of Applied Meteorology and Climatology, 46, 1053–1066.
Stewart R. E., 1992: Precipitation types in the transition region of winter storms. Bull. Amer. Meteor. Soc., 73, 287–296. Find this article online
Stewart R. E., and P. King, 1987: Freezing precipitation in winter storms. Mon. Wea. Rev., 115, 1270–1279. Find this article online.
Wandishin,M.S, M.E Baldwin, S.L. Mullen, and John V. Cortinas: 2005: Short-Range ensemble forecasts of precipitation type. Wea. And Forecasting,20,609-626.
Wetzel,S.W. and J.E. Martin 2001: An operational ingredients based methodology for forecasting midlatitude winter season precipitation. Wea. and Fore.,16,156-167.
Sivillo, S.K,J.E. Ahlquist, and Z. Toth,1997: An ensemble forecasting primer. Wea. Forecasting.,12, 809-818.
Stuart, N. A., R. H. Grumm, and M. J. Bodner, 2013: Analyzing predictability and communicating uncertainty: Lessons from the post-Groundhog Day 2009 storm and the March 2009 “megastorm.” J. Operational Meteor., 1 (16), 185–199.
Stuart,N.A and R.H . Grumm 2006: Using Wind Anomalies to Forecast East Coast Winter Storms. Wea. and Forecasting, 21,952-968.
Zsoter, Ervin, Roberto Buizza, David Richardson, 2009: “Jumpiness” of the ECMWF and Met Office EPS Control and Ensemble-Mean Forecasts. Mon. Wea. Rev., 137, 3823–3836
Figure . Total precipitation (mm) valid for the period of 1200 UTC 2 March through 1800 UTC 3 March 2014 from six European Center models runs initialized at a) 0000 UTC 25 February, b) 0000 UTC 26 February, c) 0000 UTC 27 February, d) 0000 UTC 28 February, e) 0000 UTC 1 March and f) 1200 UTC 1 March 2014. Shading indicates values in millimeters. Contours are 6.25, 12.5, 25 and 50 mm. The red dot is Louisville, KY. Return to text. Return to EFS section.
Figure . Stage-IV precipitation data (mm) showing the observed estimated precipitation across eastern North America for the period of 0000 UTC 2 March through 1200 UTC 3 March 2014. Shading shows intervals in mm and contours are 2,4,8,16,32 and 64 mm. The region was chosen to cover the main period of the event as it affected the southern Plains with sleet and a mixed event in the Ohio Valley during its evolution. Return to text.
Figure 3.The CFSR 500 hPa heights (m) and 500 hPa height anomalies in 24 hour increments from a) 1200 UTC 26 February 2014 through f) 1200 03 March 2014. Height contours every 60 m and standardized anomalies in standard deviations from normal. Return to text.
Figure 5. As in Figure 4 except for CFSR mean sea-level pressure. Return to text.
Figure 6. As in Figure 3 except 850 hPa temperatures and temperature anomalies. Return to text.
Figure 7. As in Figure 6 except for 850 hPa temperatures and anomalies focused over the Plains in 6-hour increments from a) 1800 UTC 01 March 2014 through f) 0000 UTC 3 March 2014. Return to text.
Figure 8. As in Figure 7 except for 850 hPa winds and u-wind anomalies for the period of a) 1200 02 March through f) 1800 UTC 3 March 2014. The black dot is Louisville, KY. Return to text.
Figure 9. As in Figure 2 except for EC forecasts of mean sea level pressure and pressure anomalies valid at 1200 UTC 03 March 2014. Return to text.
Figure 10. As in Figure 9 except for the EC average and spread about the mean of mean sea level pressure forecast valid at a) 0000 UTC and b) 1200 UTC 3 March 2014. Return to text.
Figure 11. As in Figure 10 except LAF for a) 500 hPa heights and b) 850 hPa temperatures valid at 1200 UTC 3 March 2014. Heights in meters every 60 m and isotherms every 2C with spread about the mean in units of each panels and based on the grey scale beneath each image. Return to text.
Figure 12. NCEP GEFS probability of 25mm or more QPF for the period of 1200 UTC 2 March 2014 through 1800 UTC 3 March 2014 from GEFS initialized at a) 0000 UTC 26 February, b) 0000 UTC 27 February, c) 0000 UTC 28 February, d) 1200 UTC 28 February, e) 0000 UTC 1 March, and f) 1200 UTC 1 March 2014. Return to text.
Figure 13. As in Figure 12 except for the ensemble mean QPF (shaded) and each members 25 mm contour. For details on heavy snow issues in the Mid-Atlantic regeion this image is zoomed in over the Mid-Atlantic region. The black dot is Harrisburg, PA. Return to text.
Figure 14. As in Figure 11a except for the mean GEFS 500 hPa heights and spread about the mean valid at 0000 UTC 3 March 2014. GEFS initialized at a) 0000 UTC 28 February, b) 0000 UTC 27 February, c) 00000 UTC 28 February, d) 1200 UTC 28 February, e) 0000 UTC 1 March and f) 1200 UTC 1 March 2014. Return to text.
Figure 15. NCEP SREF probability of 15m or more QPF for the 24 hour period ending at 1200 UTC 3 March 2014 from SREF initialized at a) 0300 UTC 28 February. b) 0900 UTC 28 February, c) 2100 UTC 28 February, d) 0300 UTC 1 March, e) 1500 UTC 1 March, and f) 2100 UTC 1 March 2014. Return to text.
Figure 16. As in Figure 15 except for the ensemble SREF mean QPF (shaded) and each members 25 mm contour. Return to text.
Figure 17. As in Figure 16 except for SREF mean pressure and the spread about the ensemble mean. Return to text.
1 Personal Communications on Social Media and the Albany MAP with videos and pictures of soundings, quarter size hail, and heavy sleet with thunder 2-3 March 2014 taken from and around Norman, OK.