• Barkmeijer, J., , R. Buizza, , and T. N. Palmer, 1999: 3D-Var Hessian singular vectors and their potential use in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 125 , 23332351.

    • Search Google Scholar
    • Export Citation
  • Barkmeijer, J., , R. Buizza, , T. N. Palmer, , K. Puri, , and J-F. Mahfouf, 2001: Tropical singular vectors computed with linearized diabatic physics. Quart. J. Roy. Meteor. Soc., 127 , 685708.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , and A. Hollingsworth, 2002: Storm prediction over Europe using the ECMWF Ensemble Prediction System. Meteor. Appl., in press.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , T. Petroliagis, , T. N. Palmer, , J. Barkmeijer, , M. Hamrud, , A. Hollingsworth, , A. Simmons, , and N. Wedi, 1998: Impact of model resolution and ensemble size on the performance of the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 124 , 19351960.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , M. Miller, , and T. N. Palmer, 1999a: Stochastic representation of model uncertainties in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 125 , 28872908.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , J. Barkmeijer, , T. N. Palmer, , and D. Richardson, 1999b: Current status and future developments of the ECMWF Ensemble Prediction System. Meteor. Appl., 6 , 114.

    • Search Google Scholar
    • Export Citation
  • Carlson, T. N., 1991: Mid-Latitude Weather Systems. Harper Collins Academic, 507 pp.

  • Douglas, R. J., 1997: Decomposition of weather forecast error using rearrangements of functions. Met Office Forecasting Res. Tech. Rep. 216, 12 pp.

    • Search Google Scholar
    • Export Citation
  • Easterling, D. R., , J. L. Evans, , P. Ya Groisman, , T. R. Karl, , K. E. Kunkel, , and P. Ambenje, 2000: Observed variability and trends in extreme weather climate events: A brief review. Bull. Amer. Meteor. Soc., 81 , 417425.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., 1994a: The Liouville equation and its potential usefulness for the prediction of forecast skill. Part I: Theory. Mon. Wea. Rev., 122 , 703713.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., . 1994b: The Liouville equation and its potential usefulness for the prediction of forecast skill. Part II: Applications. Mon. Wea. Rev., 122 , 714728.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., , J. Tribbia, , and R. M. Errico, 1995: Mesoscale predictability: An assessment through adjoint methods. Proc. ECMWF Seminar on Predictability, Reading, United Kingdom, ECMWF, 157–183.

    • Search Google Scholar
    • Export Citation
  • Epstein, E. S., 1969: A scoring system for probability forecasts of ranked categories. J. Appl. Meteor., 8 , 985987.

  • Gleeson, T. A., 1966: A causal relation for probabilities in synoptic meteorology. J. Appl. Meteor., 5 , 365368.

  • Hoffman, R. N., , Z. Liu, , J. F. Louis, , and C. Grassotti, 1995: Distortion representation of forecast errors. Mon. Wea. Rev., 123 , 27582770.

    • Search Google Scholar
    • Export Citation
  • Houtekamer, P. L., , L. Lefaivre, , J. Derome, , H. Ritchie, , and H. L. Mitchell, 1996: A system simulation approach to ensemble prediction. Mon. Wea. Rev., 124 , 12251242.

    • Search Google Scholar
    • Export Citation
  • Kocin, P. J., , and L. W. Uccellini, 1990: Snowstorms along the Northeastern Coast of the United States: 1955 to 1985. Amer. Meteor. Soc., 280 pp.

    • Search Google Scholar
    • Export Citation
  • Kunkel, K. E., , R. A. Pielke Jr., , and S. A. Changnon, 1999: Temporal fluctuations in weather and climate extremes that cause economic and human health impacts: A review. Bull. Amer. Meteor. Soc., 80 , 10771098.

    • Search Google Scholar
    • Export Citation
  • Lorenz, E. N., 1969: The predictability of a flow which possesses many scales of motion. Tellus, 29 , 505513.

  • Manobianco, M., 1989: Explosive east coast cyclogenesis over the west-central North Atlantic Ocean: A composite study derived from ECMWF operational analyses. Mon. Wea. Rev., 117 , 23652583.

    • Search Google Scholar
    • Export Citation
  • Molteni, F., , R. Buizza, , T. N. Palmer, , and T. Petroliagis, 1996: The ECMWF Ensemble Prediction System: Methodology and validation. Quart. J. Roy. Meteor. Soc., 122 , 73119.

    • Search Google Scholar
    • Export Citation
  • Mullen, S. L., , and D. P. Baumhefner, 1994: Monte Carlo simulations of explosive cyclogenesis. Mon. Wea. Rev., 122 , 15481567.

  • Mullen, S. L., , and R. Buizza, 2001: Quantitative precipitation forecasts over the United States by the ECMWF Ensemble Prediction System. Mon. Wea. Rev., 129 , 638663.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., 2000: Predicting uncertainty in forecasts of weather and climate. Rep. Prog. Phys., 63 , 71116.

  • Palmer, T. N., , F. Molteni, , R. Mureau, , R. Buizza, , P. Chapelet, , and J. Tribbia, 1993: Ensemble prediction. Proc. ECMWF Seminar on Validation of Models over Europe, Vol. I, (1992), Reading, United Kingdom, ECMWF, 21–66.

    • Search Google Scholar
    • Export Citation
  • Petroliagis, T., , R. Buizza, , A. Lanzinger, , and T. N. Palmer, 1997: Potential use of the ECMWF Ensemble Prediction System in cases of extreme weather events. Meteor. Appl., 4 , 6984.

    • Search Google Scholar
    • Export Citation
  • Sanders, F., , and J. R. Gyakum, 1980: Synoptic-dynamic climatology of the “Bomb.”. Mon. Wea. Rev., 108 , 15891606.

  • Talagrand, O., , R. Vautard, , and B. Strauss, 1999: Evaluation of probabilistic prediction systems. Proc. ECMWF Workshop on Predictability, Reading, United Kingdom, ECMWF, 1–26.

    • Search Google Scholar
    • Export Citation
  • Teixeira, J., 1999: The impact of increased boundary layer vertical resolution on the ECMWF forecast system. ECMWF Res. Dept. Tech. Memo. 268, 55 pp.

    • Search Google Scholar
    • Export Citation
  • Tiedtke, M., 1989: A comprehensive mass flux scheme for cumulus parameterization in large-scale models. Mon. Wea. Rev., 117 , 17791800.

    • Search Google Scholar
    • Export Citation
  • Tiedtke, M., . 1993: Representation of clouds in large-scale models. Mon. Wea. Rev., 121 , 30403061.

  • Toth, Z., 1997: Ensemble forecasting at NCEP and the breeding method. Mon. Wea. Rev., 125 , 32973319.

  • Toth, Z., , and E. Kalnay, 1993: Ensemble forecasting at NMC: The generation of perturbations. Bull. Amer. Meteor. Soc., 74 , 23172330.

  • View in gallery

    Observed 24-h precipitation (mm) accumulated between 1200 UTC of (a) 24th and 25th and (b) 25 and 26 Jan 2000. (Precipitation data were kindly provided by Steve Mullen). (c)–(d) Observed 24-h snowfall (mm of equivalent water) at some stations of the East Coast (these are the stations that reported snow depth on 24, 25, and 26 Jan between 80° and 70°W longitude and between 30° and 50°N latitude) accumulated over the same period. (e) Observed 2 m temperature (from the ECMWF analysis) at 1200 UTC 25th. (f) As in (e) but for 26th. Contour isolines: 1, 5, 10, 20, and 40 mm for precipitation, and every 2°C for temperature

  • View in gallery

    MSLP (a) at 0000 UTC and (b) 1200 UTC 25 Jan, and (c) at 0000 UTC and (d) 1200 UTC 26 Jan. Contour interval 5 hPa.

  • View in gallery

    MSLP forecasts valid for 1200 UTC 25 Jan. (a) Rmse of the TL319 operational high-resolution (OHR) model (solid line with diamonds), EPS control (dashed line with squares), ensemble mean (dotted line with triangles) and best perturbed-member (chain-dashed line with crosses). (b) MSLP intensity error (IE, hPa) of the TL319 operational high-resolution model (solid line with diamonds) and the EPS control (dashed line with squares). (c) As in (b) but for the position error (PE, km). (d) Number of EPS members with IE/PE smaller than 5 hPa/200 km (white pattern), with IE/PE between 5 hPa/200 km and 10 hPa/400 km (gray pattern) and with IE/PE between 10 hPa/400 km and 15 hPa/600 km (black pattern)

  • View in gallery

    As in Fig. 3 but for MSLP forecasts valid for 1200 UTC 26 Jan

  • View in gallery

    (a) MSLP verification valid at 1200 UTC 25 Jan. Other panels: 72-h forecasts started 23 Jan: (b) OHR TL319, (c) EPS control, (d) 72-h ensemble mean, (e) EPS member 36 (smallest rmse), (f) EPS member 34 (second lowest rmse), (g) EPS member 25 (lowest IE), (h) EPS member 50 (second lowest IE), and (i) EPS member 11 (third lowest IE). Contour interval is 5 hPa. In the forecast titles, rms is the forecast rmse, ie the intensity error, and pe the position error; for the TL319, no is the number of EPS perturbed-members with rmse smaller than the TL319; for the EPS control, nc is the number of EPS perturbed-members better than the control; for the EPS members, irms is the ranking position with respect to the 50 perturbed forecasts in terms of rmse, and i_ie is the ranking position in terms of IE

  • View in gallery

    As in Fig. 5 but for 48-h forecasts started 22 Jan and valid for 25 Jan

  • View in gallery

    As in Fig. 5 but for 72-h forecasts started 23 Jan and valid for 26 Jan

  • View in gallery

    As in Fig. 5 but for 48-h forecasts started 23 Jan and valid for 26 Jan

  • View in gallery

    Snowfall (left panels) and total precipitation (right panels) forecasts valid for 24-h fields accumulated between 1200 UTC 24th and 25th Jan [obs fields are shown in Figs. 1(a) and 1(c)]. (a) TL319 24-h snowfall forecast started on 24th. (b) As in (a) but for total precipitation. (c)–(d) As in (a)–(b) but for the t + 48h forecasts started 23d. (e)–(f) As in (a)–(b) but for the 72-h forecasts started 22d. Contour isolines: 1, 5, 10, 20, and 40 mm

  • View in gallery

    As in Fig. 9 but for fields accumulated between 1200 UTC 25 and 26 Jan [obs fields are shown in Figs. 1(b) and 1(d)]

  • View in gallery

    (a) Probability of “24-h precipitation in excess of 10 mm” predicted on 24th and valid from 1200 UTC 24 Jan to 1200 UTC 25 Jan. (b) As in (a) but for snowfall. (c) As in (a) but for the 48-h predictions started 23d. (d) As in (c) but for snowfall. (e) As in (a) but for the 72-h prediction started 22d. (f) As in (e) but for snowfall. Contour isolines: 2%, 10%, 30%, and 60%

  • View in gallery

    As in Fig. 11 but for the event “24-h precipitation in excess of 20 mm.”

  • View in gallery

    Observed (i.e., analysis, solid bold) and forecast by the EPS control (dash bold) and perturbed members (solid thin) position of the 0°C isotherm for 2 m temperature: (a) 24-h forecast issued 24th and analysis for 1200 UTC 25th; (b) 24-h forecast issued 25th and analysis for 1200 UTC 26th; (c) 48-h forecast issued 23d and analysis for 1200 UTC 25th; (d) 48-h forecast issued 24th and analysis for 1200 UTC 26th; (e) 72-h forecast issued 22d and analysis for 1200 UTC 25th; (f) 72-h forecast issued 23d and analysis for 1200 UTC 26 Jan 2000. The lat/long grid spacing is 2°

  • View in gallery

    (a) Probability of “24-h precipitation in excess of 10 mm” predicted on 25 Jan and valid from 1200 UTC 25 Jan to 1200 UTC 26 Jan. (b) As in (a) but for snowfall. (c) As in (a) but for the 48-h predictions started on 24th. (d) As in (c) but for snowfall. (e) As in (a) but for the 72-h prediction started on 23d. (f) As in (e) but for snowfall. Contour isolines: 2%, 10%, 30%, and 60%

  • View in gallery

    Number of MSLP perturbed-member forecasts with IE/PE smaller than 5 hPa/200 km (pattern with vertical lines), with IE/PE between 5 hPa/200 km and 10 hPa/400 km (pattern with squares) and with IE/PE between 10 hPa/400 km and 15 hPa/600 km (pattern with dots) for EPS and NOST-ensemble forecasts with different lead times: (a) for forecasts valid for 1200 UTC 25 Jan and (b) 26 Jan

  • View in gallery

    As in Fig. 5 but for 72-h NOST-ensemble forecasts started on 23 Jan and valid for 26 Jan

  • View in gallery

    MSLP 72-h forecasts started on 23 Jan and valid for 1200 UTC 26 Jan: (a) forecast difference (EPS2–CON), (b) error (EPS2–ANA), (c) difference (NOST2–CON), (d) error (NOST2–CON), (e) difference (EPS2–NOST2) and (f) error (CON–ANA). Contour interval is 5 hPa, with positive (negative) values represented by solid (dashed) lines

  • View in gallery

    As in Fig. 17 but for vertical cross section (long–height) of temperature differences averaged between 30°–60°N. Contour interval: 1° (solid (dash) for above (below) 0°C)

  • View in gallery

    Total precipitation accumulated between t + 48 h and t + 72 h for forecasts started on 23 Jan and valid for 1200 UTC 26 Jan: (a) forecast difference (EPS2–CON), (b) forecast difference (NOST2–CON) and (c) forecast difference (EPS2–NOST2). Contour interval 10 mm, [solid (dash) for positive (negative) values]

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 37 37 7
PDF Downloads 7 7 2

Prediction of the U.S. Storm of 24–26 January 2000 with the ECMWF Ensemble Prediction System

View More View Less
  • 1 European Centre for Medium-Range Weather Forecasts, Reading, United Kingdom
© Get Permissions
Full access

Abstract

Between 24 and 26 January 2000 explosive cyclogenesis along the U.S. east coast caused serious economic disruption and loss of lives. The performance of the European Centre for Medium-Range Weather Forecasts (ECMWF) high-resolution TL319 model and of the TL159 Ensemble Prediction System (EPS) in predicting the storm evolution is investigated.

The most critical time period to predict was the rapid intensification of the storm between 24 and 25 January. Single deterministic forecasts based on the TL319 model gave skillful predictions only 36 h before the event. By contrast, the EPS indicated the possibility that the storm would hit the affected region 2 days before the event, consistently enhancing the indications present in forecasts issued 3 and 4 days before the event. This suggests that the ECMWF EPS, suitably used, could be a valuable support tool for critical issues as alerts for extreme winter weather.

Sensitivity studies indicate that the interaction of initial perturbations and stochastic perturbations added to the model tendencies was a necessary ingredient to have some EPS members correctly predicting the storm.

Corresponding author address: Dr. Roberto Buizza, ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom. Email: ner@ecmwf.int

Abstract

Between 24 and 26 January 2000 explosive cyclogenesis along the U.S. east coast caused serious economic disruption and loss of lives. The performance of the European Centre for Medium-Range Weather Forecasts (ECMWF) high-resolution TL319 model and of the TL159 Ensemble Prediction System (EPS) in predicting the storm evolution is investigated.

The most critical time period to predict was the rapid intensification of the storm between 24 and 25 January. Single deterministic forecasts based on the TL319 model gave skillful predictions only 36 h before the event. By contrast, the EPS indicated the possibility that the storm would hit the affected region 2 days before the event, consistently enhancing the indications present in forecasts issued 3 and 4 days before the event. This suggests that the ECMWF EPS, suitably used, could be a valuable support tool for critical issues as alerts for extreme winter weather.

Sensitivity studies indicate that the interaction of initial perturbations and stochastic perturbations added to the model tendencies was a necessary ingredient to have some EPS members correctly predicting the storm.

Corresponding author address: Dr. Roberto Buizza, ECMWF, Shinfield Park, Reading RG2 9AX, United Kingdom. Email: ner@ecmwf.int

1. Introduction

Human activities have become increasingly vulnerable to weather: population and infrastructures continue to grow in areas exposed to extreme weather events such as flooding, strong winds, and extreme temperatures (Easterling et al. 2000; Kunkel et al. 1999). Extreme events are often associated with very energetic phenomena (e.g., hurricanes, intense winter storms) during which errors in the initial conditions (initial uncertainties) could grow very quickly and affect the forecast accuracy (Lorenz 1969; Ehrendorfer et al. 1995). Furthermore, model uncertainties due to the discrete representation of the system equations could enhance the forecast error growth. These two sources of error can lead to inconsistent deterministic forecasts started on consecutive days, thus making it difficult to issue extreme-weather warnings.

More sophisticated prediction systems capable of forecasting the time evolution of the probability density function (PDF) of forecast states should be used in these cases. The prediction of this PDF is possible theoretically and can be achieved by integrating the analogue (in numerical weather prediction) of the Fokker–Planck equation for the PDF (Gleeson 1966; Epstein 1969). Unfortunately, the Fokker–Planck equation and its perfect-model equivalent, the Liouville equation, are only tractable for very simple systems (Ehrendorfer 1994a,b), but are intractable for the currently used high-resolution weather prediction models, due to the very large number of degrees of freedom of the system (Palmer 2000).

Ensemble prediction provides a practical tool for estimating the PDF of the forecast state using multiple integrations of the model equations. Since 1992, ensemble prediction systems have become part of the operational numerical weather prediction practice at the European Centre for Medium-Range Weather Forecasts (ECMWF; Palmer et al. 1993; Molteni et al. 1996), at the National Centers for Environmental Prediction (NCEP; Toth and Kalnay 1993, 1997) and at the Canadian Meteorological Center (CMC, operational since 1995; see Houtekamer et al. 1996). Currently, both the ECMWF and the CMC ensemble systems are designed to simulate the effect of both initial and model uncertainties.

Ensemble prediction can be helpful in forecasting the risk of damage due to extreme weather events: it can be used to attach a level of confidence to a single deterministic forecast and to estimate the probability of occurrence of any event. This latter aspect is particularly important in cases of severe weather events (Petroliagis et al. 1997; Buizza and Hollingsworth 2002).

The prediction of the explosive cyclogenesis that caused serious disruption and the loss of lives on the U.S. east coast between 24 and 26 January 2000 is considered in this work. A typical case of explosive cyclogenesis (Manobianco 1989; Carlson 1991), this storm was characterized in the early stages by a surface pressure drop of about 1 hPa h−1 and by heavy snowfall over part of the east coast of the United States. This storm was not properly forecast by single deterministic forecasts issued by ECMWF and by NCEP until 36 h in advance. This type of storm occurs several times per year (about 6 storms per year, see Fig. 10.9 in Carlson 1991), with disastrous effects on coastal areas. Forecast success is case dependent and forecasts of storm intensity are frequently wrong (Mullen and Baumhefner 1994).

The rapid evolution of these systems is due to a combination of synoptic and mesoscale factors, among which strong sea surface temperature gradients play a fundamental role (Sanders and Gyakum 1980). In fact, the Pacific and Atlantic regions influenced by the Kuroshio Current and the Gulf Stream are the areas most affected by explosive cyclogenesis. Lack of data in those areas, the inherent numerical approximation of moist processes, and the fast error growth typical of explosive systems are some of the reasons for the forecast failures in predicting such events.

In this work, predictions given by the ECMWF TL159L40 Ensemble Prediction System (EPS) are compared with ECMWF deterministic high-resolution TL319L60 predictions. These two forecasting systems have a horizontal resolution of approximately 120 km and 60 km, respectively, and thus forecasts should not be expected to reveal details at scales smaller than these. This inherent limitation should be taken into account when assessing forecast performance.

The paper is organized as follows. In section 2 the configuration of the ECMWF EPS operational at the time of the storm is briefly described. The synoptic situation during the U.S. storm is discussed in section 3. The accuracy of single deterministic and ensemble forecasts are analyzed in section 4. Sensitivity experiments designed to investigate the impact on EPS perturbed forecasts of initial perturbations and stochastic perturbations added to the model tendencies are discussed in section 5. Conclusions are drawn in section 6.

2. Description of the ECMWF Ensemble Prediction System

Routine real-time execution of the ECMWF EPS started in December 1992 with a 33-member T63L19 configuration (spectral triangular truncation T63 and 19 vertical levels, Palmer et al. 1993; Molteni et al. 1996). Since 1 May 1994, the EPS has been run daily with 1200 UTC as starting time.

A major upgrade to a 51-member TL159L31 system (spectral triangular truncation T159 with linear grid) took place in 1996 (Buizza et al. 1998). This spectral resolution is equivalent to a midlatitude gridpoint spacing of about 120 km. A scheme to simulate model uncertainties due to random model error in the parameterized physical processes was introduced in 1998 (Buizza et al. 1999a). The number of vertical levels was increased from 31 to 40 in October 1999, with the extra levels in the planetary boundary layer (Teixeira 1999).

Schematically, each EPS forecast ej is generated by integrating the perturbed model equations
i1520-0493-130-6-1531-e1
where A and P′ identify the contribution to the full equation tendency of the nonparameterized and parameterized physical processes. For each grid point x = (λ, ϕ, σ) (identified by its latitude, longitude, and vertical hybrid coordinate), the perturbed parameterized tendency P′ (of each state vector component) is defined as
i1520-0493-130-6-1531-e3
where P is the unperturbed diabatic tendency, and 〈…〉D,T indicates that the same random number rj has been used for all grid points inside a D° × D° box and over T time steps. The random numbers are currently sampled uniformly in the interval [−0.5, 0.5], the same random number is used inside 10° boxes (D = 10), and the set of random numbers is updated every 6 h (T = 6) (note that random numbers do not vary with the vertical coordinate σ).
In Eq. (2) ej(t = 0) is the operational analysis at t = 0, while δej denotes the jth initial perturbation. For each day d, the initial perturbations are defined using the singular vectors that grow in the forecast range between day d and day d + 2 at initial time and the singular vectors that had grown in the past between day d − 2 and day d at final time
i1520-0493-130-6-1531-e4
where vd,d+2i (t = 0) is the ith singular vector growing between day d and d + 2 at time t = 0 (Barkmeijer et al. 1999). The coefficients αi,j and βi,j set the initial amplitude of the ensemble perturbations, and are defined by comparing the singular vectors with estimates of analysis errors (Molteni et al. 1996).

It is worth mentioning that following extensive experimentation the EPS horizontal spectral truncation was increased from TL159 to TL255 on 21 November 2000 (Buizza and Hollingsworth 2002).

3. Synoptic description

Figure 1 shows the observed total precipitation (in mm of equivalent water; Figs. 1a,b) between 1200 UTC 24 and 25 January and snowfall accumulated (Figs. 1c,d) between 1200 UTC 25 and 26 January. The observed precipitation gridpoint field has been constructed by averaging station values inside 1.25° boxes (as in Mullen and Buizza 2001: observed precipitation data were kindly provided by Dr. S. Mullen). Observed snowfall values are shown for stations archived in the ECMWF database and located along the U.S. east coast that reported snow-depth measurements at 1200 UTC 24, 25, and 26 January (snowfall is reported in mm of equivalent water and has been computed from snow-depth measurements).

Surface temperatures along the U.S. east coast are about 0°C on 25 January and below zero on 26th with a strong gradient along the coast between land and sea (Figs. 1e,f). Average precipitation fields indicate gridpoint maxima between 20 and 30 mm (Fig. 1a), with some stations in North Carolina reporting between 30 and 60 mm of rainfall between 24 and 25 January. Precipitation over land started in some locations with snow and ice on 24th and continued as snow on 25 and 26 January. The heavy snowfall had a serious impact on the East Coast: record snowfall was reported across North Carolina, with Raleigh–Durham International Airport, Raleigh, North Carolina (hereafter Raleigh) (36°N, 78°W) reporting 5 cm of snowfall in 4 h (http://www.emc.ncep.noaa.gov/mmb/research/blizz2000). A 10-cm snow depth was recorded at New York's La Guardia Airport and 28 cm of snow depth at Baltimore–Washington International Airport.

Snowfall verification is not easy since snowfall data from few stations were available. One approach could be to analyze the precipitation field in conjunction with the low-level thermal structure, and interpret the precipitation as snowfall wherever a below-freezing surface temperature is reported. If this approach is followed, then the observed precipitation field between 1200 UTC 24 and 25 January north of say 35°N (North Carolina, Virginia, Washington area; Fig. 1a) should be considered as snowfall since below-freezing temperatures are reported for this area (Fig. 1a). By contrast, all the observed precipitation between 1200 UTC 25th and 26th should be considered as snowfall since the 0°C 2-m temperature isoline runs along the East Coast (Fig. 1f). The other approach would be to consider snowfall only for the data reported as snowfall and shown in Figs. 1c,d. Note that there is a good agreement between the observed precipitation and snowfall value, and the thermal structure. This second approach is followed throughout the paper when assessing the accuracy of snowfall predictions.

Although each system has its own peculiarities (Kocin and Uccellini 1990), some of the usual features of intensity and duration common to this kind of development (Carlson 1991) are present also at this time. On 24 January 2000, a surface low pressure system starts developing to the northeast of Florida, downstream of an upper-level trough, and in phase with the jet stream axis. At upper levels (not shown), the system's evolution is characterized by a strong vorticity advection, in concert with intense cold advection upstream of the surface and warm advection downstream. During the subsequent hours the cyclone intensifies while moving along the strong temperature gradient located between the coast and the edge of the Gulf Stream. Figure 2 shows the storm position at the surface on 25 and 26 January. Between 1200 UTC 24th and 1200 UTC 25th the surface central pressure drops by about 23 hPa, from 1002 to 979 hPa. The surface pressure minimum remains between 980 and 1000 hPa during the following 18 h while the low pressure system moves northward along the coast.

4. Forecast performance

Verification focuses on mean sea level pressure (MSLP), total precipitation (defined as the sum of rainfall and snowfall) and snowfall at key times. MSLP forecasts are verified at 1200 UTC 25th and 26th and precipitation forecasts are verified against observed values accumulated between 24th and 25th and between 25 and 26 January 2000. Snowfall is predicted directly by the ECMWF cloud scheme (Tiedtke 1989, 1993). Snowfall is verified for six stations archived in the ECMWF database that reported snow depth at 1200 UTC 24, 25, and 26 January (Bagotville, Quebec, Canada; Robertval, Quebec, Canada; La Guardia Airport; Baltimore–Washington International Airport; Rochester, New York; and Raleigh; see Table 1). Snowfall forecast values have been interpolated from the four closest grid points surrounding each of the six stations. Both snowfall and total precipitation are reported in mm of equivalent water (1 mm of water is normally equivalent to approximately 1 cm of snowfall over ground).

a. Verification measures

The accuracy of deterministic forecasts of MSLP is assessed by computing root-mean-square error (rmse), intensity, position, and 24-h-tendency error (IE, PE, TE). Root-mean-square errors are computed inside a box of 30° in latitude and 40° in longitude centered on the cyclone position at the verification time. Intensity and position errors are computed by comparing the observed cyclone intensity and position with the forecast relative minimum inside a 1200-km radius from the observed cyclone (the software that identifies the cyclone position was kindly provided by Martin Leutbecher). Tendency errors are computed by comparing the observed and the forecast change in MSLP at the center of the storm over a 24-h forecast period. The rmse is a simple and widely used measure to assess forecast accuracy. The rmse is sensitive to errors in gradient or intensity, so slight displacements of features characterized by strong gradients can lead to large rmses even if forecasts are qualitatively good. More sophisticated objective measures, in which errors are partitioned into misplacement and shape errors (Hoffman et al. 1995; Douglas 1997) are still under development and are not commonly used in operational environments. The combined use of rmse, IE, PE, and TE is contrasted with a subjective verification. This approach should give a comprehensive overview of MSLP forecast accuracy. Precipitation and snowfall predictions are assessed subjectively by comparing observed and predicted fields.

b. Performance of the TL319 and the EPS control forecasts

Figure 3 shows the root-mean-square, intensity, and position errors of the TL319 and the EPS control MSLP forecasts with initial conditions from 18 (t + 168 h forecast) to the 24 (t + 24 h) January and verified at 1200 UTC 25th. The two forecasts have very similar root-mean-square and intensity errors (Figs. 3a,b), while the higher-resolution TL319 forecast has lower position errors (Fig. 3c) for all lead times apart for 24 h. Intensity errors are larger than 5 hPa for forecasts with a 48-h or a longer lead time (Fig. 3b), and position errors are 200 km or larger for all forecast times (Fig. 3c). Table 2 lists the MSLP tendency errors in predicting the observed 23 hPa day−1 pressure fall for forecasts with a lead time of 96 h or less. Table 2 shows that the TL319 have a TE larger than 9 hPa day−1 for lead times of 48 h or longer (the TE of the EPS control forecast is similar to the TE of the TL319 forecast, not shown).

Figure 4 shows the errors of the ECMWF TL319 and the EPS control MSLP forecasts verified on 26th (1200 UTC). The TL319 and the EPS control forecasts have similar root-mean-square and intensity errors (Figs. 4a,b) while the TL319 forecast has lower position errors (Fig. 4c). Compared to the forecasts with the same lead time verified on 25 January (Fig. 3b), forecasts verified on 26 January (Fig. 4b) have smaller intensity errors. The situation at the surface is easier to predict than the day before: the storm central pressure at 1200 UTC 26 January is higher than the day before (991.1 hPa compared to 984.4 hPa) and the 24-h variation in the pressure minimum is a factor of 3 smaller. Table 2 also shows that the TE of the ECMWF TL319 and the EPS control forecasts verified on 26 January are smaller than the TE of the forecasts valid for 25 January.

Figures 5 through 8 show the observed and predicted flow at the surface. The first row of Fig. 5 shows the verifying analysis at 1200 UTC 25 January and the t + 72 h MSLP forecasts given by the TL319 and the EPS control issued on 22d. Both forecasts predict a rather large area of low pressure with a smooth pressure gradient, with minimum values northeast of the observed minimum. The first row of Fig. 6 shows the corresponding t + 48 h forecasts produced the day after and still valid for 1200 UTC 25 January. The 48-h forecasts (Figs. 6b,c) predict a more localized area of low pressure than the 72-h forecasts, but they still position the pressure minimum too far off the East Coast and fail to correctly intensify the cyclonic system (Table 2).

The first row of Figs. 7 and 8 show the verifying analysis valid for 1200 UTC on 26 January and the t + 72 h and the t + 48 h MSLP forecasts given by the TL319 and the EPS control. As in the case of the forecasts valid for 1200 UTC 25th, the t + 72 h forecasts valid for 26th (Figs. 7b,c) fail to predict the location of the area of low pressure and its associated pressure gradient. The 48-h forecasts (Figs. 8b,c) predict almost correctly the position but overintensify the cyclonic development.

Figure 9 shows the TL319 24-h-accumulated forecasts of precipitation (left) and snowfall (right) issued on 24th (t + 24 h), 23d (t + 48 h), and 22d (t + 72 h), and valid for 25 January. (Since the t + 72 h and the t + 48 h EPS control forecasts are very similar to the TL319 forecast, only the latter are shown.) Over land, the t + 24 h TL319 precipitation forecast (Fig. 9a, see Fig. 1a for observations) predicts the area of precipitation almost correctly, with a small misplacement of the area of intense precipitation (40 mm) around 34°N and with no indication of observed precipitation north of 40°N (Philadelphia and New York area). Figures 9a and 9b show that over the northeastern United States precipitation is mostly predicted as snowfall, with no snowfall predicted south of 32°N (Georgia, Florida). The t + 48 h TL319 forecasts (Figs. 9c,d) are less accurate than the 24-h forecasts, with the band of precipitation not predicted far enough inland and with no indication of more than 10 mm of precipitation around 34°N (the Carolinas). The t + 72 h forecasts (Figs. 9e,f) are less accurate, with only 1–5 mm of rainfall predicted over land. The first three columns of Table 1 lists the observed snowfall and the TL319 48-h and 72-h predictions at the six selected stations. Results are in agreement with the synoptic evaluation, and confirm that deterministic predictions underestimate the observed snowfall at La Guardia, Baltimore–Washington International Airport, and Bagotville and overestimate snowfall at Robertval and Rochester.

Figure 10 shows the TL319 24-h-accumulated forecasts of precipitation (left) and snowfall (right) issued on 25th (t + 24 h), 24th (t + 48 h), and 23d (t + 72 h), and valid for 26 January. Over land, the t + 24 h TL319 precipitation forecast (Fig. 10a) fails to predict the area of more than 20 mm precipitation (Fig. 1b) around 40°–42°N (New York and Boston areas). The comparison between the maps of forecast precipitation and snowfall (Figs. 10a,b) indicates that almost all precipitation is predicted as snowfall, in agreement with the observed thermal structure (Fig. 1f). The t + 48 h TL319 forecast (Fig. 10c) keeps the band of precipitation too close to the coast and fails to produce more than 10 mm of precipitation. The comparison between the t + 48 h precipitation and snowfall predictions indicates again that almost all precipitation is predicted as snowfall. The t + 72 h forecast (Fig. 10e) fails to predict any precipitation over land where it is observed (Fig. 1b). The last three columns of Table 1 lists the observed snowfall and the TL319 48-h and 72-h predictions at the six selected stations. Results indicate a general underestimation of snowfall apart for the 72-h forecast for La Guardia.

c. Performance of the ECMWF EPS

The EPS performance in predicting MSLP is assessed by considering the number of EPS members with intensity/position errors smaller than 5 hPa/200 km, 10 hPa/400 km, and 15 hPa/600 km. Figure 3d shows this diagnostic for the EPS forecasts with initial conditions from 18th (t + 168 h forecast) to 24th (t + 24 h), and valid for 25 January (1200 UTC). Results indicate that for lead times of 120 h or longer no EPS members predict the storm with IE/PE smaller than 10 hPa/400 km, and that two EPS members predict the storm with such an accuracy for 96- and 72-h lead times. This number increases substantially (from 2 to 15) in the following 24 h. Table 2 lists some statistics of the EPS performance in terms of tendency errors. Table 2 shows that on 21st (t + 96 h) nine EPS members predict that storm intensification with a TE smaller than 5 hPa day−1 (five members with TE smaller than 2.5 hPa day−1). On 22d (t + 72 h), 3 EPS members predict the storm intensification with a TE of less than 5 hPa day−1 and 11 members predict the storm with a TE smaller than 10 hPa day−1.

Figures 5 and 6 show (second and third row) the MSLP forecasts given by the ensemble-mean and by five EPS members started on 22 (+72 h) and 23 (+48 h) January. The five EPS members are chosen to include the two members with the smallest rmse and the three members with the smallest IE. Figure 5 shows that members 25, 50, and 11 (last row) predict the storm with IE/PE of 1 hPa/1065 km, 4.9 hPa/425 km, and 5 hPa/169 km, respectively. By contrast, the TL319 and the EPS control forecasts (Fig. 5, top row) have IE/PE of 19.3 hPa/400 km and 19.3 hPa/609 km, respectively. In rmse terms, 10 EPS members have an error smaller than the TL319 and 18 members have an error smaller than the EPS control. Figure 6 shows that the 48-h forecasts are more accurate with four EPS members with IE less than 0.3 hPa (member 3, 9, 40, and 41, see middle and bottom rows of Fig. 6), among which is one (member 3) with a 78-km PE. By contrast, both the TL319 and the EPS control forecasts have IE/PE larger than 9.4 hPa/564 km (see top row of Fig. 6). For both forecast steps the ensemble-mean forecast (first panel of the second row in Figs. 5 and 6) is too smooth a field to be used to predict such an extreme event associated with a strong pressure gradient.

Figure 4d shows the number of EPS members with intensity/position errors smaller than 5 hPa/200 km, 10 hPa/400 km, and 15 hPa/600 km for the forecasts with initial conditions from 19th (t + 168 h forecast) to 25th (t + 24 h), and valid for 26 January (1200 UTC). Results show that for any lead time at least one EPS member has IE/PE smaller than 10 hPa/400 km, and that this number increases as the lead time decreases. Considering IE/PE smaller than 5 hPa/200 km, Fig. 4d shows that two or more members have such accuracy only for lead times of 72 h or less. Table 2 shows that forecasts valid for 1200 UTC of 26th more accurately predict also the storm intensification, with more than nine members with a TE smaller than 2.5 hPa day−1 for lead times of 96 h or shorter.

Figures 7 and 8 show the MSLP forecasts given by the ensemble-mean and by five EPS members (selected as discussed above) started on 23 (+72 h) and 24 (+48 h) January and valid for 1200 UTC of 26. Figure 7 (last row) shows that EPS members 22, 47, and 7 predict the storm with, respectively, IE/PE of 3.2 hPa/518 km, 3.2 hPa/189 km, and 3.7 hPa/494 km. By contrast, the TL319 and the EPS control forecasts (Fig. 7, top row) have IE/PE of, respectively, 13 hPa/541 km and 20.1 hPa/770 km. In rmse terms, 27 EPS members have an error smaller than the TL319 and 30 members had an error smaller than the EPS control. Figure 8 shows that 48-h forecasts are more accurate, with three EPS members with IE less than 0.8 hPa (members 13, 48, and 50, see middle and bottom rows of Fig. 7), among which one (member 13) with a 56-km PE. By contrast, both the TL319 and EPS control forecasts have IE larger than 6 hPa (see top row of Fig. 8). For the 72-h forecast range, the ensemble-mean forecast (first panel of the second row in Fig. 7) proves again to be too smooth a field to be used to predict extreme events, while the 48-h ensemble-mean (Fig. 8) predicts the observed cyclone with a high degree of accuracy. This is due to the fact that a large number of 48-h perturbed forecasts have very low intensity and position errors (Fig. 4d).

Consider now precipitation and snowfall forecasts. Attention is focused on lead times up to 72 h and predictions of probabilities of 24-h accumulated precipitation or snowfall exceeding predefined thresholds. Figure 11 shows the probability of “24-h precipitation in excess of 10 mm” (left) and “24-h accumulated snowfall in excess of 10 mm” (right) predicted on 24th (t + 24 h), 23d (t + 48 h), and 22d (t + 72 h) and valid for the 24 h ending at 1200 UTC of 25 January. Precipitation predictions can be compared over land with the observed field (Figs. 1a,b). Figure 11 shows that the t + 24 h forecast issued on 24th gives a 60% probability of precipitation in excess of 10 mm over the area where 10 mm is observed (see Fig. 1a). The t + 72 h forecast issued on 22d gives a 2%–10% probability of precipitation in excess of 10 mm (Fig. 11e) over the area where 10 mm is observed. The 48-h EPS indicates a 10%–30% probability of precipitation in excess of 10 mm (Fig. 11c) over the East Coast where 10 mm of precipitation is observed (Fig. 1a), with a 30%–60% probability over North Carolina (about 36°N). Forecast probabilities are even higher in the 24-h forecast (Fig. 11a). Similarly, the probabilities of snowfall in excess of 10 mm increases from zero in the t + 72 h forecast (Fig. 11f), to 2%–10% in the t + 48 h forecast (Fig. 11d) and up to 30% in the t + 24 h forecast (Fig. 11b). Figure 12 is the equivalent of Fig. 11 but for a 20 mm threshold. For this higher threshold the probability of precipitation is only slightly smaller than for 10 mm and covers a slightly smaller area over land, but snowfall probabilities are about a factor of 3-to-5 smaller. In particular, the t + 48 h forecast gives a 10% probability of snowfall in excess of 20 mm only between 34° and 36°N (the Carolinas). The signal in the EPS forecast is consistent in the consecutive forecasts, increasing from small to large values as the verification time approaches.

The fact that for equivalent thresholds, snowfall probabilities are lower than precipitation probabilities is a consequence of the fact that each EPS member has a different forecast thermal structure. Figure 13 shows the observed position of the 0°C isotherm for 2 m temperature and its position in the 72-h, 48-h, and 24-h forecasts valid for 25th (left) and 26th (right). Figure 13 shows that for the forecasts valid for 25th the observed position of the 0°C isotherm is close to the edge of the range spanned by the EPS, especially in the 48-h and the 72-h forecasts.

Figure 14 shows the probability of “24-h precipitation in excess of 10 mm” (left) and “24-h accumulated snowfall in excess of 10 mm” (right) predicted on 25th (t + 24 h), 24th (t + 48 h), and 23d (t + 72 h) and valid for the 24 h ending at 1200 UTC of 26 January. Precipitation predictions can be compared over land with the observed field (Fig. 1b). Figure 14a shows that the 24-h EPS issued on 25th gives a 60% or higher probability of precipitation in excess of 10 mm in a coastal region north of 36°N, in agreement with the observed field (Fig. 1c), but fails to predict the possibility of more than 10 mm of precipitation in a very small region centered at 35°N, 78°W (Raleigh). The 24-h EPS gives a 10%–30% probability of snowfall in excess of 10 mm for a smaller region north of 38°N (Fig. 14b) where more than 10 mm of snowfall is observed (Fig. 1d). The 48-h EPS probability forecasts issued on 24th for both precipitation and snowfall in excess of 10 mm (Figs. 14c,d) cover a larger area and have lower probability values than 24-hEPS probability forecasts. The 72-h EPS probability maps cover a small area over the northeast coast and give a 2%–10% probability of precipitation and snowfall in excess of 10 mm. The right panels of Fig. 13 show the observed and forecast positions of the 0°C isotherm for the 2-m temperature. Figure 13 shows that almost all EPS members predict the position the 0°C isotherm for the 2-m temperature too far inland compared to the observed position, which lies close to the southeastern edge of the distribution of 0°C isotherm forecast positions. Forecast probability maps of precipitation and snowfall in excess of 20 mm cover a much smaller area and have smaller values (not shown), qualitatively in agreement to the fact that 20 mm are observed only over a small region centered near 44°N (Maine, New Hampshire).

Table 3 lists the EPS snowfall predictions at the six selected stations. Note that, for all stations, the observed value lies inside the EPS predicted range for almost all forecasts. The only large difference can be detected in the 72-h forecast for Baltimore–Washington International Airport (Table 3d, 28 mm observed and 22.8 mm maximum prediction).

d. Summary of deterministic and probabilistic prediction

1) Verification time 1200 UTC 25 January—Deterministic predictions

  • The 72-h MSLP (Figs. 3 and 5, Table 2) and precipitation (Fig. 9, Table 2) forecasts fail.
  • The 48-h MSLP forecasts fail to predict the storm position and rapid intensification (Figs. 3 and 6, Table 2); precipitation and snowfall predictions give some indications of coastal impact but keep the predicted area too little inland and underestimate the precipitation amount by about a factor of 2 (Fig. 9, Table 1).
  • The 24-h MSLP TL319 and the EPS control forecasts predict the storm (200 and 360 km PE, respectively, and about 2 hPa IE; Fig. 3, Table 2); precipitation and snowfall predictions are reasonably accurate in position and amount (Fig. 9, Table 1).

2) Verification time 1200 UTC 25 January—Probabilistic predictions

  • The 72-h MSLP prediction (Figs. 3d and 5, Table 2) gives some indication of a storm approaching the East Coast (two members with IE < 10 hPa, PE < 400 km, TE < 5 hPa day−1); the 72-h prediction gives a 2%–10% probability of precipitation and snowfall in excess of 10 mm (Figs. 11 and 12, Table 3).
  • The 48-h MSLP prediction has several members with a very accurate prediction of the storm position and rapid intensification (five members with IE < 5 hPa, PE < 200 km, and TE < 2.5 hPa day−1, 16 members with IE < 10 hPa, PE < 400 km, and TE < 5 hPa day−1; see Figs. 3d and 6, Table 2); 48-h prediction gives 10%–60% (2%–30%) probabilities of precipitation and snowfall in excess of 10 mm (20 mm) over the East Coast where this amount of precipitation is observed (Figs. 11 and 12, Table 3).
  • The 24-h MSLP prediction has 14 members with IE/PE smaller than 5 hPa/200 km (Fig. 3d); the 24-h prediction gives a 60%–100% (30%–100%) probability of precipitation in excess of 10 mm (20 mm) in the region where this is observed (Figs. 11 and 12).

3) Verification time 1200 UTC 26 January—Deterministic predictions

  • The 72-h MSLP (Figs. 4 and 7, Table 2) and precipitation (Fig. 10, Table 1) forecasts fail.
  • The 48-h MSLP TL319 and the EPS control forecasts predict the storm position with a PE of about 200 km, an IE of 6 and 8 hPa, respectively, and a TE of about 3 hPa day−1 (Figs. 4 and 8, Table 2); precipitation and snowfall predictions almost correctly identify the area but underestimate the largest amounts by about a factor of 2 (Fig. 10, Table 1).
  • The 24-h MSLP TL319 and the EPS control forecasts predict the storm with a very small PE and slightly overintensify the minimum pressure (IE of about 2 hPa; Fig. 4); precipitation and snowfall predictions are rather accurate in position and amount (Fig. 10, Table 1).

4) Verification time 1200 UTC 26 January—Probabilistic predictions

  • The 72-h MSLP prediction (Figs. 4d and 7, Table 2) gives some accurate predictions of the storm moving northward along the East Coast (2 members with IE < 5 hPa, PE < 200 km, TE < 2.5 hPa day−1, seven members with IE < 10 hPa, PE < 400 km, TE < 5 hPa day−1); 72-h prediction gives a 2%–10% probability of precipitation/snowfall (Fig. 14, Table 3);
  • The 48-h MSLP prediction has many members with a very accurate prediction of the storm position and rapid intensification (8 members with IE < 5 hPa, PE < 200 km, and TE < 2.5 hPa day−1; 21 members with IE < 10 hPa, PE < 400 km, and TE < 5 hPa day−1; see Figs. 4d and 8, Table 2); 48-h prediction gives a 30%–60% probability of precipitation and snowfall in excess of 10 mm over the East Coast where more than 10 mm of precipitation is observed (Fig. 14, Table 3);
  • The 24-h MSLP predictions are very accurate (20 members have IE < 2.5 hPa, PE < 200 km, and TE < 2.5 hPa day−1; see Fig. 4d and Table 2); precipitation and snowfall predictions are very accurate in position and amount (Fig. 14).

5. Role of initial perturbations and of stochastic physics

A different set of ensemble experiments has been performed to investigate the influence of stochastic perturbations added to the tendencies due to parameterized physical processes (Buizza et al. 1999a). The results of this sensitivity analysis should be considered as a documentation of the relative impact of stochastic perturbations compared to the initial perturbations. The reader is referred to published literature (e.g., Buizza et al. 1999a; Mullen and Buizza 2001) for more complete sensitivity studies based on larger sets of cases. These experiments can also give some indications regarding the sensitivity of forecast errors to moist processes during this case.

The new set of ensembles (NOST) has been run at the same resolution as the operational ensemble but without stochastic perturbations for starting dates from 19 to 25 January. Considering MSLP predictions, Fig. 15 shows the number of ensemble members with intensity/position errors smaller than 5 hPa/200 km, 10 hPa/400 km, and 15 hPa/400 km for the NOST and the operational EPS (run with stochastic physics). The ensembles run without the stochastic forcing have a smaller number of members with small IE/PE for all lead times apart from the 48-h forecasts started on 24th and verified on 26 January (Fig. 15b). Results indicate that the NOST ensembles are less able to correctly intensify the cyclone, especially for forecast times longer than 48 h.

An example of the positive impact of stochastic physics is given by the comparison of the +72 h forecast started on 23d and valid for 26th. Figure 16 shows selected NOST ensemble members to be compared with the selected EPS members shown in Fig. 7. It is interesting to note that member number 18 and member number 2 have the best rmse for both types of ensembles, and that stochastic physics reduces the rmse of one of them and the IE for both forecasts. The comparison of the three NOST and EPS members with the smallest IE suggests that the cyclone intensity and position is more accurate in the EPS than in the NOST forecast.

Precipitation and snowfall predictions given by the NOST ensemble have been compared with the respective predictions given by the operational EPS. Little differences can be identified between the NOST ensemble (not shown) and the EPS 48-h probability maps for the events “precipitation/snowfall larger than 10 mm and 20 mm.” The difference is more evident in the 72-h forecasts: over the East Coast, the EPS probabilities extend further inland than the NOST probabilities (not shown), in better agreement with the observations. In Table 4 the average observed value (among the six selected stations) is contrasted with the EPS and the NOST-ensemble median, 75th and 95th percentiles and maximum predicted values. Results indicate that the EPS predicted values are about 10% larger than the NOST values.

Three 72-h forecasts valid for 1200 UTC on 26 January are compared, to document in more detail, the impact of initial perturbations and stochastic physics. The selected forecasts are the EPS control forecast (CON), the EPS member 2 (EPS2), which is the EPS member with the second-lowest rmse and the third-lowest IE (see Fig. 7f), and member 2 run without stochastic physics (NOST2, see Fig. 16f). Figures 7 and 16 show that EPS2 has a lower rmse, a lower IE but a larger PE than NOST2 (rmse: 3.4 hPa vs 4.1 hPa; IE: 3.4 hPa vs 9.6 hPa; PE: 558 km vs 183 km). The 72-h control forecast shows two weak minima (Fig. 7c), both wrongly located, one at about (55°N, 40°W) and one at about (40°N, 65°W). By contrast, EPS2 shows (Fig. 7f) a unique, more correctly positioned minimum.

Figure 17 shows the differences among forecasts (EPS2–CON), (NOST2–CON), and (EPS2–NOST2) and the errors of the three forecasts in terms of MSLP. The left panels of Fig. 17 show that (EPS2–CON) and (NOST2–CON) are rather similar, and that (EPS2–NOST2) is smaller than (EPS2–CON) and (NOST2–CON). This indicates that stochastic physics has a smaller impact than the initial perturbation added to generate the perturbed initial condition of EPS member number 2. Figure 17e shows that stochastic physics intensifies the cyclone in the 72-h forecast. The comparison of the error fields (Fig. 17, right panels) shows that both EPS2 and NOST2 have smaller errors than the control. The analysis of the evolution of (EPS2–CON) with forecast time (not shown) indicates that the difference at t + 72 h starts at t + 24 h over northern Florida (at about 30°N, 80°W) and propagates northeastward along the coast line.

Figure 18 is analogous to Fig. 17 but shows a vertical cross section (longitude vs pressure) of temperature differences averaged in latitude between 30° and 60°N. The left panels of Fig. 18 confirm that stochastic physics has a smaller impact than the initial perturbation and that both EPS2 and NOST2 have smaller errors than the control forecast. Stochastic physics induces some cooling in the lower troposphere around 70°W (Figs. 18a,c) in correspondence of the difference in MSLP (Figs. 17a,c). The initial perturbation added to member 2 induces a warming of the whole air column in correspondence with the cyclone evolution (Fig. 18c), and this reduces substantially the temperature errors (Fig. 18, right panels).

Figure 19 is analogous to Figs. 17 and 18 but shows differences between 24-h accumulated precipitation predictions valid for 26 January. Figure 19 shows that EPS2 and NOST2 both have more intense precipitation over land than the control forecast, and that stochastic physics induces a shift of the precipitation band over sea but has a very small effect over land.

6. Conclusions

Explosive cyclogenesis over the east coast of United States is a relatively rare event (Kocin and Uccellini 1990) occurring about 6 times per year (Carlson 1991), that is still rather difficult to predict, with skillful prediction alternating with forecast failures (Mullen and Baumhefner 1994). One of these severe winter weather events affected the U.S. east coast between 24 and 26 January 2000, when MSLP dropped about 23 hPa between 1200 UTC of 24th and 25th and snowfall between 25th and 26th paralyzed part of the U.S. east coast. The performance of the ECMWF TL319L60 high-resolution model and of the TL159L40 Ensemble Prediction System during the U.S. storm of 25 and 26 January 2000 has been discussed. Attention has been focused on mean sea level pressure (MSLP), snowfall, and total precipitation predictions.

The most critical period to predict coincided with the rapid development of the storm between 24 and 25 January. For this period, single (TL319 and EPS control) deterministic forecasts issued 48-h and 72-h before the event failed to intensify the storm, wrongly predicted the area affected by precipitation over land and underestimated the overall amount of precipitation by about a factor of 2. Shorter, 24-h deterministic forecasts also failed to correctly predict the storm position and intensification, but at least gave a more accurate precipitation prediction. By contrast, 72-h EPS forecasts gave some indication of the possibility of the storm developments and predict a 2%–10% probability of precipitation and snowfall in excess of 10 mm in a region where 10 mm values were observed. The 48-h EPS forecasts were more accurate, with 16 members predicting the storm with intensity error IE < 10 hPa, position error PE < 400 km, and tendency error TE < 5 hPa day−1. The 48-h EPS prediction gave a higher, 10%–60% probability of precipitation and snowfall in excess of 10 mm and a 2%–30% probability of values in excess of 20 mm for regions where these amounts were observed. The 24-h EPS predictions had 14 members with intensity and position errors of IE < 5 hPa and PE < 200 km and gave high probabilities of precipitation and snowfall in excess of 10 and 20 mm in the area where these amounts were observed.

Both deterministic and probabilistic predictions of the storm development between 25 and 26 were more accurate.

The comparison of the single (TL319 and EPS control) deterministic and the EPS probabilistic forecasts indicates that the EPS gave some indications of the developments of the storm 24 to 48 h before the single deterministic forecasts. The consistency of EPS forecasts issued on consecutive days made ensemble probabilistic forecasts particularly useful.

One of the key issues to be addressed when promoting the use of ensemble forecast products is how forecasters should use ensemble-generated products to assess the risk of severe weather disruption. Despite the fact that guidelines can only be drawn after a statistical analysis of a large sample of cases, this work gives some indications of possible ensemble products. Results suggest that for this type of severe winter weather event, EPS probability maps of precipitation and snowfall, maps showing the position of the 0°C isoline of the 2-m temperature field combined with synoptic charts of MSLP of all ensemble members can be used to assess the severity of a forecast situation. Results also indicate that the ensemble-mean field is too smooth a field to be used to predict severe weather events associated with a strong pressure gradient.

Confidence in the quality of the EPS can be consolidated only through statistical analyses of the performance over long periods (Talagrand et al. 1999). This study has been limited to a detailed discussion of one case only of extreme winter weather. Thus, only limited conclusions on the usefulness of the EPS as a tool to predict the risk of extreme weather can be drawn. The reader is referred to published literature (Buizza et al. 1999b; Barkmeijer et al. 2001; Mullen and Buizza 2001; Buizza and Hollingsworth 2002), for more detailed studies of the performance of the EPS.

Results from sensitivity experiments have shown that stochastic physics has, in general, a smaller impact than initial perturbations on the EPS perturbed forecasts and that the combination of initial perturbations and stochastic physics improves the quality of the EPS performance. These results confirm the positive impact of the stochastic scheme on the EPS performance reported in Buizza et al. (1999a) and Mullen and Buizza (2001). These sensitivity results suggest that forecast errors in the prediction of the 24–26 January U.S. storm were very sensitive to moist processes.

Acknowledgments

The authors would like to thank Martin Leutbecher, who developed a first version of the code to identify MSLP minimum values and Steve Mullen for providing observed precipitation data for the U.S. storm case study. Anders Persson, Tony Hollingsworth, and Tim Palmer are acknowledged for their useful comments to an early version of this manuscript. The authors would also like to thank the editor Dr. D. P. Jorgensen and three anonymous reviewers whose comments and suggestions helped improving the quality of the manuscript.

REFERENCES

  • Barkmeijer, J., , R. Buizza, , and T. N. Palmer, 1999: 3D-Var Hessian singular vectors and their potential use in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 125 , 23332351.

    • Search Google Scholar
    • Export Citation
  • Barkmeijer, J., , R. Buizza, , T. N. Palmer, , K. Puri, , and J-F. Mahfouf, 2001: Tropical singular vectors computed with linearized diabatic physics. Quart. J. Roy. Meteor. Soc., 127 , 685708.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , and A. Hollingsworth, 2002: Storm prediction over Europe using the ECMWF Ensemble Prediction System. Meteor. Appl., in press.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , T. Petroliagis, , T. N. Palmer, , J. Barkmeijer, , M. Hamrud, , A. Hollingsworth, , A. Simmons, , and N. Wedi, 1998: Impact of model resolution and ensemble size on the performance of the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 124 , 19351960.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , M. Miller, , and T. N. Palmer, 1999a: Stochastic representation of model uncertainties in the ECMWF Ensemble Prediction System. Quart. J. Roy. Meteor. Soc., 125 , 28872908.

    • Search Google Scholar
    • Export Citation
  • Buizza, R., , J. Barkmeijer, , T. N. Palmer, , and D. Richardson, 1999b: Current status and future developments of the ECMWF Ensemble Prediction System. Meteor. Appl., 6 , 114.

    • Search Google Scholar
    • Export Citation
  • Carlson, T. N., 1991: Mid-Latitude Weather Systems. Harper Collins Academic, 507 pp.

  • Douglas, R. J., 1997: Decomposition of weather forecast error using rearrangements of functions. Met Office Forecasting Res. Tech. Rep. 216, 12 pp.

    • Search Google Scholar
    • Export Citation
  • Easterling, D. R., , J. L. Evans, , P. Ya Groisman, , T. R. Karl, , K. E. Kunkel, , and P. Ambenje, 2000: Observed variability and trends in extreme weather climate events: A brief review. Bull. Amer. Meteor. Soc., 81 , 417425.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., 1994a: The Liouville equation and its potential usefulness for the prediction of forecast skill. Part I: Theory. Mon. Wea. Rev., 122 , 703713.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., . 1994b: The Liouville equation and its potential usefulness for the prediction of forecast skill. Part II: Applications. Mon. Wea. Rev., 122 , 714728.

    • Search Google Scholar
    • Export Citation
  • Ehrendorfer, M., , J. Tribbia, , and R. M. Errico, 1995: Mesoscale predictability: An assessment through adjoint methods. Proc. ECMWF Seminar on Predictability, Reading, United Kingdom, ECMWF, 157–183.

    • Search Google Scholar
    • Export Citation
  • Epstein, E. S., 1969: A scoring system for probability forecasts of ranked categories. J. Appl. Meteor., 8 , 985987.

  • Gleeson, T. A., 1966: A causal relation for probabilities in synoptic meteorology. J. Appl. Meteor., 5 , 365368.

  • Hoffman, R. N., , Z. Liu, , J. F. Louis, , and C. Grassotti, 1995: Distortion representation of forecast errors. Mon. Wea. Rev., 123 , 27582770.

    • Search Google Scholar
    • Export Citation
  • Houtekamer, P. L., , L. Lefaivre, , J. Derome, , H. Ritchie, , and H. L. Mitchell, 1996: A system simulation approach to ensemble prediction. Mon. Wea. Rev., 124 , 12251242.

    • Search Google Scholar
    • Export Citation
  • Kocin, P. J., , and L. W. Uccellini, 1990: Snowstorms along the Northeastern Coast of the United States: 1955 to 1985. Amer. Meteor. Soc., 280 pp.

    • Search Google Scholar
    • Export Citation
  • Kunkel, K. E., , R. A. Pielke Jr., , and S. A. Changnon, 1999: Temporal fluctuations in weather and climate extremes that cause economic and human health impacts: A review. Bull. Amer. Meteor. Soc., 80 , 10771098.

    • Search Google Scholar
    • Export Citation
  • Lorenz, E. N., 1969: The predictability of a flow which possesses many scales of motion. Tellus, 29 , 505513.

  • Manobianco, M., 1989: Explosive east coast cyclogenesis over the west-central North Atlantic Ocean: A composite study derived from ECMWF operational analyses. Mon. Wea. Rev., 117 , 23652583.

    • Search Google Scholar
    • Export Citation
  • Molteni, F., , R. Buizza, , T. N. Palmer, , and T. Petroliagis, 1996: The ECMWF Ensemble Prediction System: Methodology and validation. Quart. J. Roy. Meteor. Soc., 122 , 73119.

    • Search Google Scholar
    • Export Citation
  • Mullen, S. L., , and D. P. Baumhefner, 1994: Monte Carlo simulations of explosive cyclogenesis. Mon. Wea. Rev., 122 , 15481567.

  • Mullen, S. L., , and R. Buizza, 2001: Quantitative precipitation forecasts over the United States by the ECMWF Ensemble Prediction System. Mon. Wea. Rev., 129 , 638663.

    • Search Google Scholar
    • Export Citation
  • Palmer, T. N., 2000: Predicting uncertainty in forecasts of weather and climate. Rep. Prog. Phys., 63 , 71116.

  • Palmer, T. N., , F. Molteni, , R. Mureau, , R. Buizza, , P. Chapelet, , and J. Tribbia, 1993: Ensemble prediction. Proc. ECMWF Seminar on Validation of Models over Europe, Vol. I, (1992), Reading, United Kingdom, ECMWF, 21–66.

    • Search Google Scholar
    • Export Citation
  • Petroliagis, T., , R. Buizza, , A. Lanzinger, , and T. N. Palmer, 1997: Potential use of the ECMWF Ensemble Prediction System in cases of extreme weather events. Meteor. Appl., 4 , 6984.

    • Search Google Scholar
    • Export Citation
  • Sanders, F., , and J. R. Gyakum, 1980: Synoptic-dynamic climatology of the “Bomb.”. Mon. Wea. Rev., 108 , 15891606.

  • Talagrand, O., , R. Vautard, , and B. Strauss, 1999: Evaluation of probabilistic prediction systems. Proc. ECMWF Workshop on Predictability, Reading, United Kingdom, ECMWF, 1–26.

    • Search Google Scholar
    • Export Citation
  • Teixeira, J., 1999: The impact of increased boundary layer vertical resolution on the ECMWF forecast system. ECMWF Res. Dept. Tech. Memo. 268, 55 pp.

    • Search Google Scholar
    • Export Citation
  • Tiedtke, M., 1989: A comprehensive mass flux scheme for cumulus parameterization in large-scale models. Mon. Wea. Rev., 117 , 17791800.

    • Search Google Scholar
    • Export Citation
  • Tiedtke, M., . 1993: Representation of clouds in large-scale models. Mon. Wea. Rev., 121 , 30403061.

  • Toth, Z., 1997: Ensemble forecasting at NCEP and the breeding method. Mon. Wea. Rev., 125 , 32973319.

  • Toth, Z., , and E. Kalnay, 1993: Ensemble forecasting at NMC: The generation of perturbations. Bull. Amer. Meteor. Soc., 74 , 23172330.

Fig. 1.
Fig. 1.

Observed 24-h precipitation (mm) accumulated between 1200 UTC of (a) 24th and 25th and (b) 25 and 26 Jan 2000. (Precipitation data were kindly provided by Steve Mullen). (c)–(d) Observed 24-h snowfall (mm of equivalent water) at some stations of the East Coast (these are the stations that reported snow depth on 24, 25, and 26 Jan between 80° and 70°W longitude and between 30° and 50°N latitude) accumulated over the same period. (e) Observed 2 m temperature (from the ECMWF analysis) at 1200 UTC 25th. (f) As in (e) but for 26th. Contour isolines: 1, 5, 10, 20, and 40 mm for precipitation, and every 2°C for temperature

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 2.
Fig. 2.

MSLP (a) at 0000 UTC and (b) 1200 UTC 25 Jan, and (c) at 0000 UTC and (d) 1200 UTC 26 Jan. Contour interval 5 hPa.

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 3.
Fig. 3.

MSLP forecasts valid for 1200 UTC 25 Jan. (a) Rmse of the TL319 operational high-resolution (OHR) model (solid line with diamonds), EPS control (dashed line with squares), ensemble mean (dotted line with triangles) and best perturbed-member (chain-dashed line with crosses). (b) MSLP intensity error (IE, hPa) of the TL319 operational high-resolution model (solid line with diamonds) and the EPS control (dashed line with squares). (c) As in (b) but for the position error (PE, km). (d) Number of EPS members with IE/PE smaller than 5 hPa/200 km (white pattern), with IE/PE between 5 hPa/200 km and 10 hPa/400 km (gray pattern) and with IE/PE between 10 hPa/400 km and 15 hPa/600 km (black pattern)

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 4.
Fig. 4.

As in Fig. 3 but for MSLP forecasts valid for 1200 UTC 26 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 5.
Fig. 5.

(a) MSLP verification valid at 1200 UTC 25 Jan. Other panels: 72-h forecasts started 23 Jan: (b) OHR TL319, (c) EPS control, (d) 72-h ensemble mean, (e) EPS member 36 (smallest rmse), (f) EPS member 34 (second lowest rmse), (g) EPS member 25 (lowest IE), (h) EPS member 50 (second lowest IE), and (i) EPS member 11 (third lowest IE). Contour interval is 5 hPa. In the forecast titles, rms is the forecast rmse, ie the intensity error, and pe the position error; for the TL319, no is the number of EPS perturbed-members with rmse smaller than the TL319; for the EPS control, nc is the number of EPS perturbed-members better than the control; for the EPS members, irms is the ranking position with respect to the 50 perturbed forecasts in terms of rmse, and i_ie is the ranking position in terms of IE

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 6.
Fig. 6.

As in Fig. 5 but for 48-h forecasts started 22 Jan and valid for 25 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 7.
Fig. 7.

As in Fig. 5 but for 72-h forecasts started 23 Jan and valid for 26 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 8.
Fig. 8.

As in Fig. 5 but for 48-h forecasts started 23 Jan and valid for 26 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 9.
Fig. 9.

Snowfall (left panels) and total precipitation (right panels) forecasts valid for 24-h fields accumulated between 1200 UTC 24th and 25th Jan [obs fields are shown in Figs. 1(a) and 1(c)]. (a) TL319 24-h snowfall forecast started on 24th. (b) As in (a) but for total precipitation. (c)–(d) As in (a)–(b) but for the t + 48h forecasts started 23d. (e)–(f) As in (a)–(b) but for the 72-h forecasts started 22d. Contour isolines: 1, 5, 10, 20, and 40 mm

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 10.
Fig. 10.

As in Fig. 9 but for fields accumulated between 1200 UTC 25 and 26 Jan [obs fields are shown in Figs. 1(b) and 1(d)]

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 11.
Fig. 11.

(a) Probability of “24-h precipitation in excess of 10 mm” predicted on 24th and valid from 1200 UTC 24 Jan to 1200 UTC 25 Jan. (b) As in (a) but for snowfall. (c) As in (a) but for the 48-h predictions started 23d. (d) As in (c) but for snowfall. (e) As in (a) but for the 72-h prediction started 22d. (f) As in (e) but for snowfall. Contour isolines: 2%, 10%, 30%, and 60%

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 12.
Fig. 12.

As in Fig. 11 but for the event “24-h precipitation in excess of 20 mm.”

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 13.
Fig. 13.

Observed (i.e., analysis, solid bold) and forecast by the EPS control (dash bold) and perturbed members (solid thin) position of the 0°C isotherm for 2 m temperature: (a) 24-h forecast issued 24th and analysis for 1200 UTC 25th; (b) 24-h forecast issued 25th and analysis for 1200 UTC 26th; (c) 48-h forecast issued 23d and analysis for 1200 UTC 25th; (d) 48-h forecast issued 24th and analysis for 1200 UTC 26th; (e) 72-h forecast issued 22d and analysis for 1200 UTC 25th; (f) 72-h forecast issued 23d and analysis for 1200 UTC 26 Jan 2000. The lat/long grid spacing is 2°

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 14.
Fig. 14.

(a) Probability of “24-h precipitation in excess of 10 mm” predicted on 25 Jan and valid from 1200 UTC 25 Jan to 1200 UTC 26 Jan. (b) As in (a) but for snowfall. (c) As in (a) but for the 48-h predictions started on 24th. (d) As in (c) but for snowfall. (e) As in (a) but for the 72-h prediction started on 23d. (f) As in (e) but for snowfall. Contour isolines: 2%, 10%, 30%, and 60%

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 15.
Fig. 15.

Number of MSLP perturbed-member forecasts with IE/PE smaller than 5 hPa/200 km (pattern with vertical lines), with IE/PE between 5 hPa/200 km and 10 hPa/400 km (pattern with squares) and with IE/PE between 10 hPa/400 km and 15 hPa/600 km (pattern with dots) for EPS and NOST-ensemble forecasts with different lead times: (a) for forecasts valid for 1200 UTC 25 Jan and (b) 26 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 16.
Fig. 16.

As in Fig. 5 but for 72-h NOST-ensemble forecasts started on 23 Jan and valid for 26 Jan

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 17.
Fig. 17.

MSLP 72-h forecasts started on 23 Jan and valid for 1200 UTC 26 Jan: (a) forecast difference (EPS2–CON), (b) error (EPS2–ANA), (c) difference (NOST2–CON), (d) error (NOST2–CON), (e) difference (EPS2–NOST2) and (f) error (CON–ANA). Contour interval is 5 hPa, with positive (negative) values represented by solid (dashed) lines

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 18.
Fig. 18.

As in Fig. 17 but for vertical cross section (long–height) of temperature differences averaged between 30°–60°N. Contour interval: 1° (solid (dash) for above (below) 0°C)

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Fig. 19.
Fig. 19.

Total precipitation accumulated between t + 48 h and t + 72 h for forecasts started on 23 Jan and valid for 1200 UTC 26 Jan: (a) forecast difference (EPS2–CON), (b) forecast difference (NOST2–CON) and (c) forecast difference (EPS2–NOST2). Contour interval 10 mm, [solid (dash) for positive (negative) values]

Citation: Monthly Weather Review 130, 6; 10.1175/1520-0493(2002)130<1531:POTUSS>2.0.CO;2

Table 1.

Stations used to verify 24-h accumulated snowfall predictions by the TL319 operational model. The observed value for Raleigh is known only for the period from 1200 UTC 25 Jan to 1200 UTC 26 Jan. Snowfall is reported in mm of equivalent water (1 mm of water is normally equivalent to approx 1 cm of snowfall over ground)

Table 1.
Table 2.

Observed and forecast 24-h change (tendency, hPa day−1) in the minimum MSLP between 1200 UTC 24 Jan and 25 Jan (rows 1–3) and MSLP between 1200 UTC 25 Jan and 26 Jan (rows 4–6). The first three columns list the initial date and the two consecutive forecast steps, column 4 lists the obs change, column 5 lists the error TE in the TL319 tendency forecast, columns 6–8 lists the number of EPS members with TE smaller than 5 hPa d−1 (TE smaller than 2.5 hPa d−1 in brackets), with TE between 5 and 10 hPa d−1 and with TE larger than 10 hPa d−1

Table 2.
Table 3.

EPS t + 24 h, t + 48 h, and t + 72 h snowfall forecasts for (a) Bagotville, (b) Robertval, (c) La Guardia Airport, (d) Baltimore–Washington International Airport, (e) Rochester, and (f) Raleigh (24-h accumulated values are expressed in mm of equivalent water). The obs value for Raleigh is known only for the period from 1200 UTC 25 Jan to 1200 UTC 26 Jan

Table 3.
Table 4.

The 15-station avg obs value (column 1) and avg forecast statistics for the EPS (column 2–5) and for the NOST-ensemble (columns 6–9) (24-h accumulated snowfall values are expressed in mm of equivalent water)

Table 4.
Save