• Anthes, R. A., , Kuo Y-H. , , and Gyakum J. R. , 1983: Numerical simulations of a case of explosive cyclogenesis. Mon. Wea. Rev., 111 , 11741188.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Black, T. L., 1994: The new NMC mesoscale Eta Model: Description and forecast examples. Wea. Forecasting, 9 , 265278.

  • Bosart, L. F., 1981: The Presidents' Day snowstorm of 18–19 February 1979: A subsynoptic-scale event. Mon. Wea. Rev., 109 , 15421566.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bosart, L. F., , and Lin S. C. , 1984: A diagnostic analysis of the Presidents' Day storm of February 1979. Mon. Wea. Rev., 112 , 21482177.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Buizza, R., , and Chessa P. , 2002: Prediction of the U.S. storm of 24– 26 January 2000 with the ECMWF Ensemble Prediction System. Mon. Wea. Rev., 130 , 15311551.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Colle, B. A., , and Mass C. F. , 2001: Evaluation of the timing and strength of MM5 and Eta surface trough passages over the eastern Pacific. Wea. Forecasting, 16 , 553572.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Colle, B. A., , Westrick K. J. , , and Mass C. F. , 1999: Evaluation of the MM5 and Eta-10 precipitation forecasts over the Pacific Northwest during the cool season. Wea. Forecasting, 14 , 137154.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Duchon, C. E., 1979: Lanczos filtering in one and two dimensions. J. Appl. Meteor., 18 , 10161022.

  • Gelaro, R., , Reynolds C. A. , , Langland R. H. , , and Rohaly G. D. , 2000:: A predictability study using geostationary satellite wind observations during NORPEX. Mon. Wea. Rev., 128 , 37893807.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grimit, E. P., , and Mass C. F. , 2002: Initial results of a mesoscale short-range ensemble forecasting system over the Pacific Northwest. Wea. Forecasting, 17 , 192205.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grumm, R. H., , and Siebers A. L. , 1989: Systematic surface cyclone errors in NMC's nested grid model November 1988–January 1989. Wea. Forecasting, 4 , 246252.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gyakum, J. R., 1983a: On the evolution of the QE II storm. I: synoptic aspects. Mon. Wea. Rev., 111 , 11371155.

  • Gyakum, J. R., 1983b: On the evolution of the QE II storm. II: Dynamic and thermodynamic structure. Mon. Wea. Rev., 111 , 11561173.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Klinker, E., , Rabier F. , , and Gelaro R. , 1998: Estimation of key analysis errors using the adjoint technique. Quart. J. Roy. Meteor. Soc., 124 , 19091933.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kuo, Y-H., , and Reed R. J. , 1988: Numerical simulation of an explosively deepening cyclone in the eastern Pacific. Mon. Wea. Rev., 116 , 20812105.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Langland, R. H., and Coauthors, 1999: The North Pacific Experiment (NORPEX-98): Targeted observations for improved North American weather forecasts. Bull. Amer. Meteor. Soc., 80 , 13631384.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Langland, R. H., , Shapiro M. A. , , and Gelaro R. , 2002: Initial condition sensitivity and error growth in forecasts of the 25 January 2000 East Coast snowstorm. Mon. Wea. Rev., 130 , 957974.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mass, C. F., , Ovens D. , , Westrick K. , , and Colle B. A. , 2002: Does increasing horizontal resolution produce better forecasts? The results of two years of real-time numerical weather prediction in the Pacific Northwest. Bull. Amer. Meteor. Soc., 83 , 407430.

    • Search Google Scholar
    • Export Citation
  • Miguez-Macho, G., , and Paegle J. , 2000: Sensitivity of a global forecast model to initialization with reanalysis datasets. Mon. Wea. Rev., 128 , 38793889.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mullen, S. L., 1994: An estimate of systematic error and uncertainty in surface cyclone analysis over the North Pacific Ocean: Some forecasting implications. Wea. Forecasting, 9 , 221227.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • NCDC, 2001: Storm Data. Vol. 43, No. 12, 106 pp. [Available from W. Angel, Ed., National Oceanographic and Atmospheric Administration, National Climatic Data Center, Asheville, NC 28801-2733.].

    • Search Google Scholar
    • Export Citation
  • NCDC, 2002: Storm Data. Vol. 44, No. 2, 90 pp. [Available from W. Angel, Ed., National Oceanographic and Atmospheric Administration, National Climatic Data Center, Asheville, NC 28801- 2733.].

    • Search Google Scholar
    • Export Citation
  • Nelsen, J. A., 1999: The Eta data assimilation system. WR Tech. Attachment 99-14, 6 pp. [Available from National Weather Service Western Region, P.O. Box 11188, Salt Lake City, UT 84147.].

    • Search Google Scholar
    • Export Citation
  • Oravec, R. J., , and Grumm R. H. , 1993: The prediction of rapidly deepening cyclones by NMC's nested grid model: Winter 1989– autumn 1991. Wea. Forecasting, 8 , 248270.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rabier, F., , Klinker E. , , Courtier P. , , and Hollingsworth A. , 1996: Sensitivity of forecast errors to initial conditions. Quart. J. Roy. Meteor. Soc., 122 , 121150.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Reed, R. J., , and Albright M. D. , 1986: A case study of explosive cyclogenesis in the eastern Pacific. Mon. Wea. Rev., 114 , 22972319.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Black T. , , Collins W. , , Manikin G. , , Mesinger F. , , Parrish D. , , and DiMego G. , 2000: Changes to the NCEP Meso Eta Analysis and Forecast System: Assimilation of satellite radiances and increase in resolution. NWS Technical Procedures Bulletin 473, 10 pp. [Available online at http://wwwt.emc.ncep.noaa.gov/ mmb/mmbpll/eta22tpb/; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Black T. , , Ferrier B. , , Lin Y. , , Parrish D. , , and DiMego G. , 2001a: Changes to the NCEP Meso Eta Analysis and Forecast System: Increase in resolution, new cloud microphysics, modified precipitation assimilation, modified 3DVAR analysis. NWS Technical Procedures Bulletin 488, 15 pp. [Available online at www.emc.ncep.noaa.gov/mmb/mmbpll/eta12tpb; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Ek M. , , Lin Y. , , Mitchell K. , , Parrish D. , , and DiMego G. , 2001b:: Changes to the NCEP Meso Eta Analysis and Forecast System: Assimilation of observed precipitation, upgrades to land-surface physics, modified 3DVAR analysis. NWS Technical Procedures Bulletin 479, 20 pp. [Available online at www.emc.ncep.noaa.gov/mmb/mmbpll/spring2001/tpb/; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

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

  • Shapiro, M., , and Thorpe A. , cited 2002: Program overview: The observing-system research and predictability experiment, THORPEX. [Available online at www.mmm.ncar.edu/uswrp/thorpex/thorpex_plan13.pdf.].

    • Search Google Scholar
    • Export Citation
  • Smith, B. B., , and Mullen S. L. , 1993: An evaluation of sea level cyclone forecasts produced by NMC's nested-grid model and global spectral model. Wea. Forecasting, 8 , 3756.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Steenburgh, W. J., , and Mass C. F. , 1996: Interaction of an intense extratropical cyclone with coastal orography. Mon. Wea. Rev., 124 , 13291352.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Szunyogh, I., , Kalnay E. , , Morss R. E. , , Majumdar S. J. , , Etherton B. J. , , and Bishop C. H. , 2000: The effect of targeted dropsonde observations during the 1999 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 128 , 35203527.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Szunyogh, I., , Toth Z. , , Zimin A. V. , , Majumdar S. J. , , and Persson A. , 2002: Propagation of the effect of targeted observations: The 2000 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 130 , 11441165.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., 1986: The possible influence of upstream upper- level baroclinic processes on the development of the QEII storm. Mon. Wea. Rev., 114 , 10191027.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., , Kocin P. J. , , Petersen R. A. , , Wash C. H. , , and Brill K. F. , 1984:: The Presidents' Day cyclone of 18–19 February 1979: Synoptic overview and analysis of the subtropical jet streak influencing the pre-cyclogenetic period. Mon. Wea. Rev., 112 , 3155.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., , Keyser D. , , Brill K. F. , , and Wash C. H. , 1985: Presidents' Day cyclone of 18–19 February 1979: Influence of upstream trough amplification and associated tropopause folding on rapid cyclogenesis. Mon. Wea. Rev., 113 , 962988.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Westrick, K., , Mass C. , , and Colle B. , 1999: The limitations of the WSR-88D radar network for quantative precipitation measurement over the coastal western United States. Bull. Amer. Meteor. Soc., 80 , 22892298.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, F., , Snyder C. , , and Rotunno R. , 2002: Mesoscale predictability of the “surprise” snowstorm of 24–25 January 2000. Mon. Wea. Rev., 130 , 16171632.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • View in gallery

    NCEP Eta Model 24-h (solid) and 48-h (dashed) forecast of SLP (mb) verifying at 0000 UTC 3 Mar 1999. The corresponding GOES-10 infrared satellite image is also shown. The analyzed surface low location and estimated central pressure from subjective surface analyses are indicated by the larger “L.”

  • View in gallery

    Map of buoy (e.g., 46001) and coastal (e.g., DESW1) station locations used in the study

  • View in gallery

    Time series of SLP errors defined as the forecast SLP minus the observed SLP (mb) at the station. The average error at each station for each forecast time is plotted as a solid line and plus/minus 2 times the std dev is shown as dashed lines. Arrows and numbers 1, 2, and 3 refer to specific events discussed in the text. (a) The 24-h forecast errors at CMAN station Tatoosh Island, WA (TTIW1), (b) 24-h forecast errors at CMAN station Cape Arago, OR (CARO3), (c) 48-h forecast errors at Tatoosh Island, and (d) 48-h forecast errors at Cape Arago

  • View in gallery

    Histograms of 48-h forecast SLP error (mb) at Tatoosh Island, for the four winter seasons. The average error and std dev are provided for each season

  • View in gallery

    Mean 500-mb heights and rms calculated from time-filtered (2–10 days) 500-mb heights over the North Pacific and coastal regions for the (a) 1999/2000 winter season, (b) 2000/01 winter season, (c) 2001/02 winter season, and (d) 2002/03 winter season

  • View in gallery

    Histograms of cyclone central pressure and position errors for the events listed in Table 4. The average errors and mean absolute errors are also indicated: (a) 24-h and (b) 48-h forecasts

  • View in gallery

    The 48-h forecasts of SLP (solid) and 00-h AVN analysis of SLP (dashed) overlaid on the infrared satellite image valid 0000 UTC 8 Feb 2002 by the (a) AVN, (b) Eta, (c) UKMO,  and (d) NOGAPS models

  • View in gallery

    As in Fig. 7, except for 24-h forecasts of SLP valid 0000 UTC 8 Feb

  • View in gallery

    Initial condition differences between the UKMO and the Eta analyses at 0000 UTC 7 Feb 2002. (a) The 500-mb heights from the UKMO (solid) and Eta (dashed) and the difference (ETA − UKMO; shaded). The differences are shaded every 10 m starting at 30 m. In addition, the trough lines from the UKMO heights and Eta heights are shown with solid and dashed black lines, respectively. (b) The 850-mb temperature (K) from the UKMO (solid) and Eta (dashed), with the positive differences (UKMO − Eta) in shades of gray every 0.5°C and negative differences contoured in gray every 0.5°C

  • View in gallery

    Initial SLP fields and available surface reports for the (a) AVN, (b) Eta, (c) UKMO, and (d) NOGAPS model analyses for 0000 UTC 7 Feb 2002. The corresponding GOES-W infrared satellite image is also shown

  • View in gallery

    The 24-h forecasts of SLP (solid) and the AVN analysis of SLP (dashed) overlaid on the infrared satellite image valid at 0000 UTC 8 Feb 2002 by the (a) MM5-AVN, (b) MM5- Eta, (c) MM5-UKMO, and (d) MM5-NOGAPS

  • View in gallery

    The 48-h forecasts of SLP (solid) and 00-h AVN SLP analysis (dashed) overlaid on the infrared satellite image valid at 0000 UTC 14 Dec 2001 by the (a) AVN, (b) Eta, (c) UKMO, and (d) NOGAPS models

  • View in gallery

    (a) Forecast error of cyclone central pressure (mb) and (b) cyclone position (km) as a function of forecast lead time verifying at 0000 UTC 14 Dec 2001 from the AVN, CMC, Eta, NOGAPS, and UKMO models. The verifying position and central pressure is obtained from subjective analysis

  • View in gallery

    Initial condition differences for the 48-h forecasts initialized at 0000 UTC 12 Dec. (a) The 500-mb heights (m) produced by the UKMO (solid) and the NOGAPS (dashed) models and the difference (UKMO − NOGAPS; shaded). The differences are shaded in gray every 10 m starting at 30 m. (b) The 850-mb temperature (K) from the UKMO (solid) and NOGAPS (dashed) models and the positive differences (UKMO − NOGAPS) in shades of gray every 0.5°C and negative differences contoured in gray every 0.5°C

  • View in gallery

    SLP analyses and ship and buoy surface reports at 0000 UTC 12 Dec produced by the (a) AVN, (b) NOGAPS, and (c) UKMO models. Location of incipient surface lows obtained from the NCEP marine surface analyses are indicated by the two larger “L”s at 37.5°N, 178°W and 39°N, 177°E

All Time Past Year Past 30 Days
Abstract Views 0 0 0
Full Text Views 210 209 10
PDF Downloads 20 20 2

Major Numerical Forecast Failures over the Northeast Pacific

View More View Less
  • 1 Department of Atmospheric Sciences, University of Washington, Seattle, Washington
© Get Permissions
Full access

Abstract

Strong North Pacific storms that impact the North American west coast are sometimes poorly predicted in the short term (up to 48 h) by operational models, with cyclone position errors of hundreds of kilometers and central pressure errors of tens of millibars. These major numerical forecast failures still occur despite continuing improvements in modeling and data assimilation. In this paper, the frequency and intensity of sea level pressure errors at buoy and coastal locations are documented by comparing the National Centers for Environmental Prediction (NCEP) Eta Model forecasts to observations and through case studies of two poorly forecast cyclones from the 2001/02 winter season.

Using data from October 1999 through March 2003 at coastal and offshore sites along the west coast of North America, it was found that large forecast errors (48-h sea level pressure errors greater than 10 mb) by the Eta Model occur 10–15 times each winter, and extremely large errors (48-h errors greater than 15 mb) occur 3–4 times per winter. Such substantial forecast errors are often associated with large position errors of surface low pressure centers. For example, storms associated with large 48-h forecast errors greater than 10 mb at nearshore and coastal sites had average forecast position errors of 453 km and mean absolute central pressure errors of 7.5 mb.

To illustrate the nature of such large forecast errors, two major cyclones that were poorly predicted by several operational models are examined. The 7–8 February 2002 storm was a compact, but powerful, cyclone that struck western Oregon with strong winds, injured four people, and produced extensive damage and power outages. The 24-h numerical forecasts for this event were poor and had a variety of solutions. Two operational models forecast lows of sufficient depth, but displaced them more than 150 km to the east or southeast of the verifying position. Three other operational models did not produce a low at all but only predicted weak troughs. The comparison of the initial conditions of the various models revealed large differences, with the more accurate models starting with sharper, more intense features. The 13–14 December 2001 storm developed rapidly offshore of British Columbia, Canada, and brought extensive rain, winds, and snow to the mountains along the west coast. The 48-h forecasts of sea level pressure by five different operational numerical models had very large errors, with cyclone position errors greater than 400 km and central pressure errors on the order of 10 mb. Differences among the initial conditions of these operational models were smaller than in the February case. Comparison of the initial conditions to surface observations revealed potentially significant errors in the vicinity of the incipient cyclone.

Corresponding author address: Dr. Lynn McMurdie, Dept. of Atmospheric Sciences, Box 351640, University of Washington, Seattle, WA 98195. Email: mcmurdie@atmos.washington.edu

Abstract

Strong North Pacific storms that impact the North American west coast are sometimes poorly predicted in the short term (up to 48 h) by operational models, with cyclone position errors of hundreds of kilometers and central pressure errors of tens of millibars. These major numerical forecast failures still occur despite continuing improvements in modeling and data assimilation. In this paper, the frequency and intensity of sea level pressure errors at buoy and coastal locations are documented by comparing the National Centers for Environmental Prediction (NCEP) Eta Model forecasts to observations and through case studies of two poorly forecast cyclones from the 2001/02 winter season.

Using data from October 1999 through March 2003 at coastal and offshore sites along the west coast of North America, it was found that large forecast errors (48-h sea level pressure errors greater than 10 mb) by the Eta Model occur 10–15 times each winter, and extremely large errors (48-h errors greater than 15 mb) occur 3–4 times per winter. Such substantial forecast errors are often associated with large position errors of surface low pressure centers. For example, storms associated with large 48-h forecast errors greater than 10 mb at nearshore and coastal sites had average forecast position errors of 453 km and mean absolute central pressure errors of 7.5 mb.

To illustrate the nature of such large forecast errors, two major cyclones that were poorly predicted by several operational models are examined. The 7–8 February 2002 storm was a compact, but powerful, cyclone that struck western Oregon with strong winds, injured four people, and produced extensive damage and power outages. The 24-h numerical forecasts for this event were poor and had a variety of solutions. Two operational models forecast lows of sufficient depth, but displaced them more than 150 km to the east or southeast of the verifying position. Three other operational models did not produce a low at all but only predicted weak troughs. The comparison of the initial conditions of the various models revealed large differences, with the more accurate models starting with sharper, more intense features. The 13–14 December 2001 storm developed rapidly offshore of British Columbia, Canada, and brought extensive rain, winds, and snow to the mountains along the west coast. The 48-h forecasts of sea level pressure by five different operational numerical models had very large errors, with cyclone position errors greater than 400 km and central pressure errors on the order of 10 mb. Differences among the initial conditions of these operational models were smaller than in the February case. Comparison of the initial conditions to surface observations revealed potentially significant errors in the vicinity of the incipient cyclone.

Corresponding author address: Dr. Lynn McMurdie, Dept. of Atmospheric Sciences, Box 351640, University of Washington, Seattle, WA 98195. Email: mcmurdie@atmos.washington.edu

1. Introduction

Intense extratropical cyclones over the North Pacific often impact the west coast of North America with strong winds, heavy precipitation, and major societal impacts. Several times each year, short-term (0 to 48 h) model forecasts of strong North Pacific cyclones are seriously deficient, with position errors measured in the hundreds of kilometers and sea level pressure errors exceeding 15 mb. For example, the National Centers for Environmental Prediction (NCEP) Eta Model 48-h forecast valid 0000 UTC 3 March 1999 predicted a low center that was 24 mb too weak and positioned 500 km southeast of the observed location (Fig. 1). The 48-h forecast from the NCEP Aviation (AVN) model was better but still possessed large errors, with the low center's central pressure being 13 mb too weak and displaced 200 km to the south of the observed position (not shown). Based primarily on the Eta forecasts, the National Weather Service predicted snow over the lowlands of western Washington as the forecast low center passed to the south of the region (J. Albrecht 2000, personal communication). In reality, the low moved to the north, producing considerable damage over Washington and Oregon due to strong winds and wave damage. In addition, a decision was made to tow out to sea a grounded cargo ship, the New Carissa. However, extremely high seas from the storm snapped the towline, causing the ship to drift ashore and leak large quantities of fuel oil into an environmentally sensitive coastal area. In this paper, the frequency and severity of such short- term forecasting failures along the west coast during the four most recent winters are documented. In addition, we examine two cyclones from the 2001/02 winter season that were plagued by poor numerical guidance.

Prior to the mid-1980s, operational numerical models often failed to predict rapid cyclogenesis (Sanders and Gyakum 1980). Notable forecast failures of major cyclone developments include the QEII storm of September 1978 (Gyakum 1983a, b; Anthes et al. 1983; Uccellini 1986), the President's Day Storm of February 1979 (Bosart 1981; Bosart and Lin 1984; Uccellini et al. 1984, 1985), and the eastern Pacific storm of November 1981, which was associated with a Limited-Area Fine Mesh (LFM) model error of 55 mb at 48 h (Reed and Albright 1986; Kuo and Reed 1988). Several authors in the next decade evaluated the performance of the NCEP Nested Grid Model (NGM) and the NCEP AVN model and found that they typically underpredicted the intensity of oceanic cyclones and often had large displacement errors (Grumm and Siebers 1989; Smith and Mullen 1993; Oravec and Grumm 1993; Mullen 1994). The Inaugural Day windstorm of January 1993, which produced $130 million in damage and left 750 000 homes without power, was underpredicted by the NCEP Eta Model by about 20 mb and had a displacement error of 150 km at 36 h (Steenburgh and Mass 1996).

Problematic forecasts of East Coast snowstorms (such as 24–26 January 2000) have brought renewed attention to the failures of numerical models to predict high-impact midlatitude cyclones. Langland et al. (2002) demonstrated through adjoint sensitivity methods using the Navy Operational Global Atmospheric Prediction System (NOGAPS) that rapid growth of small but critical initial condition errors over the eastern Pacific and western/central United States caused the 72-h forecast failures of the January 2000 storm. Buizza and Chessa (2002) showed that the European Centre for Medium- Range Weather Forecasts (ECMWF) ensemble system indicated the possibility of a major snowstorm at a greater lead time than the ECMWF deterministic forecast. Zhang et al. (2002) performed several fifth-generation Pennsylvania State University–National Center for Atmospheric Research Mesoscale Model (MM5) experiments and found that the mesoscale distribution of precipitation was significantly altered by small changes in initial conditions. Despite all the attention this storm has received, the short-term forecast (24–48 h) errors of the surface low (5–6 mb, 50–200 km for the 24-h forecasts) were actually quite modest by Pacific standards; indeed, they were far less than those of the March 1999 east Pacific case noted above or the west coast cases discussed later in this paper.

A central question regarding short-term forecast failures over the west coast concerns the role and magnitudes of initial condition errors. As described in this paper, there are often large initial condition errors over the eastern Pacific that cause significant short-term forecast failures. On the other hand, several authors have demonstrated that small errors in highly sensitive regions can grow into large forecast errors (Rabier et al. 1996; Klinker et al. 1998; Gelaro et al. 2000, among others). In particular, Gelaro et al. (2000) found large forecast improvements over western North America by reducing analysis error in highly sensitive regions determined from singular vector analysis. Klinker et al. (1998) found similar results using adjoint sensitivity techniques. In contrast, Miguez-Macho and Paegle (2000) found that large-scale errors of the initial state might play a more prominent role in forecast errors than suggested by singular vector analysis.

It has long been recognized that the predictability of cyclones over North America is adversely affected by the lack of observations over the Pacific (Shapiro and Thorpe 2002). To address this issue, several recent field experiments have been conducted over the North Pacific, such as the North Pacific Experiment (NORPEX) and the National Oceanic and Atmospheric Administration (NOAA) Winter Storms Reconnaissance (WSR) Program. Their primary focus has been to improve downstream weather forecasts over land by securing targeted observations over data-sparse and/or sensitive regions over the North Pacific Ocean. The results have been mixed, although forecasts were improved in a majority of cases (Langland et al. 1999; Gelaro et al. 2000; Szunyogh et al. 2000, 2002). The Winter Storms Reconnaissance Program of targeted deployment of dropsondes from aircraft has continued over the North Pacific (Szunyogh et al. 2002) during recent winter seasons.

In recognition of the importance of oceanic regions in the predictability (or lack of predictability) of synoptic-scale systems over North America and elsewhere, a multifaceted field program, The Hemispheric Observing System Research and Predictability Experiment (THORPEX; Shapiro and Thorpe 2002) is being planned. This research and field program will examine predictability issues on many time scales (24 h through 2 weeks) and for several regions of the world. Among other things, the field program will deploy several innovative observing systems, including new generation satellite products. Goals of the project include identifying factors that limit predictability of high-impact synoptic events and assessing the relative contribution of initial condition errors and model errors on predictability.

This paper continues the study of oceanic predictability by documenting the frequency and severity of major recent forecast failures over the northeast Pacific, including the U.S. West Coast. In addition, we examine two high-impact events that suffered from poor short- term numerical forecasts. Our goal is to demonstrate the severity of the numerical forecast problem on the west coast and to set the stage for future investigations, including THORPEX. The paper is organized as follows: In section 2, results of model verification using west coast buoys are given, and in section 3 two case studies are presented. In section 4, the results are discussed, followed by the summary and conclusions.

2. Eta Model verification using buoy and coastal observations

a. Datasets and methods

In this portion of the study, sea level pressure observations from offshore and coastal NOAA buoys over the eastern Pacific and the NOAA Coastal Marine Automated Network (CMAN) stations along the west coast were compared to model forecasts. Oceanic and coastal stations were chosen since we were interested in the quality of the forecasts before the storms made landfall. The locations of the buoy and CMAN sites used in the study are shown in Fig. 2. The station data were quality controlled to remove unrealistic values, such as surface pressures greater than 1050 mb and less than 950 mb (see Colle et al. 1999).

The period of study is from 1 October 1999 through 1 April 2003, which encompasses the four most recent winter seasons. The NCEP Eta Model (Black 1994) was used for the study because of the long record of Eta verification statistics collected as part of the University of Washington (UW) MM5 verification system (Mass et al. 2002). During the above period, the Eta had several upgrades, including changes in resolution and data assimilation. At the beginning of the period, the Eta was run at 32 km and 45 levels. In September 2000 the Eta Model was upgraded to 22 km and 50 levels, and since late November 2001 it has been run at 12 km and 60 levels. Throughout this period, the Eta was initialized using a three-dimensional variational data assimilation (3DVAR) system [the Eta Data Assimilation System (EDAS); see Nelsen (1999)]. Several changes to the data assimilation system, including additional data sources, occurred over the study period. For additional information about these changes see Rogers et al. (2000, 2001a,b). The grids used for the comparison with observations are the Eta 104 grids, where the model output is interpolated to a 90-km horizontal grid and a 50-mb (before 2002) or a 25-mb (after 2002) vertical resolution.

The 24- and 48-h Eta forecasts (initialized at 0000 and 1200 UTC each day) of sea level pressure were compared to the observed sea level pressure at each buoy and CMAN location. To do so, the Eta forecasts were interpolated to observation sites using a bilinear interpolation scheme. Time series of mean sea level pressure (SLP) error at each station, defined as forecast sea level pressure minus observed, were calculated over the study period. In addition, the average error, the mean absolute error, and the standard deviation of the errors using winter season (October–March) data only for all years were calculated at each observation site.

For each station, dates and times when the absolute value of the SLP error was greater than 2 times the winter season standard deviation (2*SD) were identified. A large error “event” was defined by two or more adjacent stations1 having errors greater than 2*SD for two or more consecutive forecasts for the same storm. Initial analyses and forecast maps were examined to determine whether each event was associated with a surface low, high, or frontal trough. For the events associated with surface lows, the forecast errors of the central sea level pressure and low positions were determined using the Eta initial analyses as verification.

b. Results

In Fig. 3, time series of SLP errors at the CMAN stations at Tatoosh Island (48.4°N, 124.7°W; station ID TTIW1 in Fig. 2), at the northwest tip of the Olympic Peninsula along the Washington coast, and at Cape Arago (43.3°N, 124.4°W; station ID CARO3 in Fig. 2), along the southwest Oregon coast, are shown for 24- and 48-h forecasts. In addition to the time series, the average error (solid line) and 2*SD error level (dashed lines) for all four winters (October–March) are provided at both locations. At both stations, the winter seasons are characterized by larger errors and error variability, compared to the other times of the year. The 48-h forecast errors are significantly larger than the 24-h errors. For example, the standard deviation of the 48-h forecast errors at Tatoosh Island is 4.1 mb, while the 24-h standard deviation is 2.6 mb.

It is clear from Fig. 3 that there are particular forecasts with exceptionally large errors. Three such events are highlighted in Fig. 3 with the numbers 1, 2, and 3. Event 1 (13–14 February 2000) was a rapidly developing storm that formed off the California–Oregon border. For this event, both the 24- and 48-h Eta forecast errors were larger than 10 mb at central Oregon's Cape Arago, which was near the path of the surface low. For event 2 (9–11 January 2001) the Eta Model forecast a major storm to move into British Columbia, Canada, whereas, in fact, that storm weakened and a second low formed offshore of California and made landfall north of San Francisco. As a result, the Tatoosh Island forecast error was large and negative, while at Cape Arago it was large and positive. In event 3 (21–23 February 2002) the Eta Model forecast a weak low to move across Washington State and a surface high to build along the coast. Instead, the surface low stalled offshore for 48 h longer than predicted, bringing significant rainfall and minor flooding to the region.

To quantify the frequency of large forecast errors, the number of 24- and 48-h forecasts for which the sea level pressure errors were greater than 2 times and 3 times the wintertime standard deviation (2*SD, 3*SD) was determined for the coastal and offshore observing sites shown in Fig. 2. They are presented in Tables 1 and 2. For either forecast projection, the number of forecasts with errors greater than 2*SD at a station ranged from approximately 10 to 25 forecasts, and the number of forecasts with errors greater than 3*SD ranged from 2 to 10 forecasts over a winter season. An error greater than 2*SD typically exceeded 5–6 mb for 24-h forecasts and 8–10 mb for 48-h forecasts. An error of 3*SD typically exceeded 7–10 mb for 24-h forecasts and 10–16 mb for 48-h forecasts.

There are geographic differences in the errors, with lower errors to the south. The standard deviation of the 48-h errors from the buoy and coastal stations north of 46°N range from 4.2 to 5.0 mb, while south of 46°N it varies from 2.9 to 3.6 mb (not shown). This pattern is apparent because wintertime surface lows typically track north of about 46°N, and errors in position or central pressure will be largest near the cyclone tracks.

Table 3 presents the average number of large errors for the seven sites2 in Tables 1 and 2, with complete records for all four winters, and provides information on the interannual variability of such large errors. Specifically, the average number of 24-h errors greater than two standard deviations decreased from 23.7 yr−1 in 1999/2000 to 13.1 yr−1 in 2000/01. In 2001/02 the average number of large errors climbed to 18.3 and then decreased to 9.0 mb in 2002/03. The 24-h three standard deviation errors followed a similar evolution. The 48- h error statistics were also similar in their interannual variability, although there were more occurrences, as would be expected for a longer forecast lead time.

To further illustrate the variation in interannual error, Fig. 4 provides histograms of 48-h sea level pressure errors for the four winter seasons at Tatoosh Island. The 2000/01 and 2002/03 winter seasons had fewer than five 48-h forecasts with absolute errors greater than 10 mb, compared to the 1999/2000 and 2001/02 winter seasons, each of which had more than 10 such events. The standard deviation of forecast errors was larger for the 1999/ 2000 and 2001/02 seasons compared to the other two seasons. Similar interannual histogram variations were found at Cape Arago and other offshore buoy stations (not shown). The average errors for each year were positive at Tatoosh Island (and at all the other stations; not shown), indicating that the Eta Model tends to have a positive forecast bias (sea level pressure too high) on average.

An important question is whether the interannual variation in large forecast errors reflects changes in storm activity or in model skill. To examine storm activity, the mean 500-mb heights and root-mean-square (rms) of time-filtered3 500-mb heights (which retain periods within the 2–10-day band) are plotted in Fig. 5 for each winter season. The 500-mb height data were obtained from the NCEP reanalysis grids of daily averaged 500- mb height. Large values of rms indicate high variability of the 500-mb heights, or storminess. These plots show that there is considerable interannual variability in storminess. The 2002/03 season stands out as the year with the strongest west coast 500-mb ridge and significantly less storminess along the west coast and throughout the northeast Pacific (Fig. 5d). The 2002/03 season also had fewer large forecast errors at Tatoosh Island (Figs. 3 and 4) and other stations (Tables 1 and 2). The 2001/02 season, a winter with a larger number of large forecast errors, had the flattest ridge and the most storminess in the northeastern Pacific than the other seasons (Fig. 5c). On the other hand, the 1999/2000 season had considerably less storminess than the 2001/02 season and yet experienced a greater number of large forecast errors than 2001/02 (Fig. 5a; Tables 1 and 2). This suggests that the Eta forecasts have become more skillful over the 4-yr study period.

In Table 4, the number of major error events (defined as period when the SLP errors were greater than 2*SD for two or more consecutive 48-h forecasts at two or more adjacent stations) for each winter season is shown. An improving trend is apparent, ranging from 21 large error events in 1999/2000 to 16 in 2001/02 and 2002/ 03. A majority of the forecast error events were associated with closed lows at the surface. For those lows associated with big errors listed in Table 4, histograms of the distribution of forecast errors of the cyclone central pressure and position for 24- and 48-h forecast lead times are shown in Fig. 6. The average cyclone central pressure 24-h forecast error for the big error events is 2.4 mb (mean absolute error is 4.1 mb), but there were about 10 forecasts with absolute errors as large as 10 mb. The average cyclone position error was 260 km, but there were 15 forecasts with errors greater than 400 km. The 48-h forecast errors were significantly larger: the average central pressure error was 3.4 mb (absolute error was 7.5 mb), and there were 20 forecasts with central pressure errors greater than 10 mb. The average cyclone position error was 453 km, and 19 forecasts had cyclone position errors of 600 km or greater. Clearly, one of the most significant problems associated with Eta forecast errors is the misplaced position of cyclones.

3. Case studies

In this section, the nature of the northeast Pacific prediction problem is illustrated by two poorly forecast cyclones from the 2001/02 winter season. As shown below, poor initialization over the Pacific strongly affected the forecasts of the 7–8 February 2002 case and possibly contributed to the forecast errors of the 13–14 December 2001 case.

a. The 7–8 February 2002 cyclone

During 7–8 February 2002, a compact but powerful cyclone struck western Oregon with wind gusts exceeding 70 kt (35 m s−1) along the southern Oregon coast and over the central Willamette Valley. These winds produced over $6 million in damage, the loss of power to 80 000 homes, and four injuries (NCDC 2002). In Eugene, Oregon, over 100 trees were downed, some 3 ft in diameter, and three cars were crushed by fallen trees (Eugene Register-Guard, 8 February 2002). The storm surprised National Weather Service forecasters, who did not mention the potential for moderate or strong winds in the morning zone forecasts on 7 February.

The storm developed south of a deep, vertically stacked low over the Gulf of Alaska. The surface low associated with the storm was very weak as it moved slowly across the eastern Pacific, roughly along 40°N, until it was just offshore of the California–Oregon border, where it deepened rapidly to 996 mb. With large pressure and isallobaric gradients, the storm produced very strong, damaging winds during the afternoon and evening of 7 February.

Forecasts from four operational models [the NCEP AVN and Eta Models, the U. K. Met Office (UKMO) Unified Model, and the Fleet Numerical Meteorological and Oceanographic Center NOGAPS] are shown in Figs. 7 and 8. The resolution of the computational and distributed grids and the objective analysis scheme used for each model at the time of this storm are given in Table 5. The 48-h forecasts of sea level pressure (solid) from the AVN, Eta, UKMO, and NOGAPS models valid at 0000 UTC 8 February and the corresponding satellite image and AVN analysis (dashed) are plotted in Fig. 7. This is approximately when the storm made landfall and the strongest winds were observed. The 48-h forecasts for this system were extremely poor by all models. The UKMO Unified Model was the only model that depicted a closed low, but its central pressure was only 1009 mb (whereas the low verified at 996 mb) and was located well east of the actual cyclone position. The AVN model [and the Canadian Meteorological Centre (CMC) Global Environmental Multiscale (GEM) model, not shown] had weak troughing over eastern Oregon and offshore of northern California, while the Eta had a weak low well offshore of northern California. The NOGAPS forecast had a ridge across eastern Oregon and Washington instead of a low.

The 24-h forecast of sea level pressure (solid) from the same four models (initialized 0000 UTC 7 February) and the corresponding satellite image and AVN analysis (dashed) are plotted in Fig. 8. Both the AVN and UKMO forecasts produced surface lows almost as deep as the AVN analysis, but displaced the lows to the southeast or east. The 24-h forecasts of the Eta, NOGAPS (Fig. 8), and CMC (not shown) models were far worse. None of these models produced a closed low, rather they developed weak offshore troughs with slack pressure gradients across western Oregon.

Since the 24-h forecast by the UKMO model was best overall for this storm, the UKMO initial conditions and the initial conditions of the other models were compared to see if coherent differences were apparent. Figure 9 shows the differences between the UKMO and Eta initial conditions (both regridded to the MM5 36-km grid) for 0000 UTC 7 February 2002 (these are the initial conditions for the 24-h forecasts shown in Fig. 8). In Fig. 9a, the UKMO (solid) and Eta (dashed) 500-mb heights and the difference between the UKMO and Eta heights (shaded) are shown. The UKMO 500-mb heights depict a trough farther east than the Eta's, and the UKMO heights are up to 50 m lower than the Eta heights over a broad region east of the UKMO trough. The UKMO 850-mb temperatures (Fig. 9b) are much warmer (up to 5°C) than the Eta's immediately east of the 850-mb low and slightly cooler than the Eta's just west of the low. Similar temperature differences are found on the 1000-mb surface (not shown). Finally, the sea level pressure initializations for four operational forecasts and the available ship and buoy reports at 0000 UTC 7 February are shown in Fig. 10. One is struck by the substantial differences among the models regarding the depth and position of the surface low that developed into the storm in question. At the surface, the UKMO surface low is in the correct location (35°N, 146°W; determined from subjective surface analysis and verified by two ship reports) and has a central pressure of 1008 mb, while the Eta surface low is too far west and has a central pressure of only 1012 mb. The AVN and NOGAPS lows are better positioned, but their low centers are not deep enough. The UKMO initial conditions at 500 and 850 mb were also compared to regridded AVN, CMC, and NOGAPS initial conditions (not shown). The UKMO initial conditions depict sharper, more intense, and better-positioned features for the incipient storm than these other models.

In order to confirm the role of initial condition errors in producing the observed forecast errors, the MM5 was run with the initial conditions from the five operational models (AVN, CMC, Eta, NOGAPS, and UKMO) over the UW operational 36-km MM5 domain (Mass et al. 2002), which extends over the northeast Pacific. In Fig. 11, the 24-h sea level pressure forecasts from the MM5- AVN, MM5-Eta, MM5-NOGAPS, and MM5-UKMO are shown. A comparison with the 24-h operational model forecasts (Fig. 8) reveals that the MM5 runs forced by the operational model initial conditions and boundary conditions were quite similar, but not identical, to the parent runs. As with the operational model results, the MM5-UKMO had the best forecast: the surface low is displaced about 100 km to the southeast of the observed location, and the forecast central pressure (999 mb) is close to the observed (996 mb). The MM5- AVN, the next most skillful member, placed the low south of the observed location. The MM5-NOGAPS and MM5-CMC (not shown) developed a weak low of only 1010 mb at the California–Oregon border, with very weak pressure gradients across Oregon. Similar to the operational Eta run, the MM5-Eta did not develop a surface low. In short, the MM5 runs with initial conditions depicting a strong incipient storm at the correct location (such as the UKMO and, to a lesser extent, the AVN initial conditions) had better forecasts than the model runs using initial conditions with a weak depiction of the incipient storm. Thus, for this event it appears that large initial condition errors played a dominant role in producing the large forecast errors.

b. The 13–14 December 2001 cyclone

The 13–14 December 2001 cyclone was one of many storms that occurred during a very active period early in the 2001/02 winter season. It developed rapidly offshore of British Columbia and by 0000 UTC 14 December 2001 a surface low pressure center of 980 mb reached the northwest coast. The storm brought strong winds over the Oregon coast and western Washington, with a significant amount of snow falling over the Washington and Oregon Cascades [over 30 cm of snow in many locations; see NCDC (2001)]. As described below, 48-h forecasts of this event were very poor, with large position errors by all operational models.

The 48-h sea level pressure forecasts of several models (AVN, Eta, NOGAPS, and UKMO) valid at 0000 UTC 14 December 2001 are shown in Fig. 12. The AVN analysis at this time (dashed lines) and the coincident infrared satellite image are also presented. Each of these forecasts (as well as the CMC forecast that is not shown) had large but different position errors of several hundred kilometers. Central pressure errors were also substantial, with the UKMO forecast of 986 mb most closely approximating the analyzed central pressure (981 mb). In contrast to the February case, the 24-h forecasts valid at 0000 UTC 14 December 2001 displayed relatively small errors (not shown), with central pressure errors ranging from 5 to 8 mb, and the forecast cyclone position errors ranging from only 50 to 200 km.

The forecast errors of cyclone central pressure and position at 0000 UTC 14 December 2001 as a function of forecast lead time by the AVN, CMC, Eta, NOGAPS, and UKMO models are given in Fig. 13. The position and central pressure of the cyclone were determined from subjective surface analysis. For the 48-h forecasts, there is considerable variability in the central pressure errors, ranging from 6 mb for the UKMO model to 14 mb for the AVN. Short lead-time forecasts have relatively small position errors, and longer lead-time forecasts have large position errors. For example, at 48 h the position errors range from 390 km for the NOGAPS to 510 km for the Eta and UKMO models. The spread of the magnitude of the position error among the members is relatively small, particularly at 48 h.

Since the 48-h forecasts had large errors, we examined whether there were obvious differences in the initializations for the 48-h forecasts initialized at 0000 UTC 12 December 2001. The incipient storm was located near the date line at this time and therefore only the AVN, UKMO, and NOGAPS initial conditions (for which we had global grids) can be used. The largest initial condition differences were found between the UKMO and NOGAPS models and are given in Fig. 14. The initial condition differences between the AVN and the other two models are slightly smaller but qualitatively similar (not shown). In Fig. 14a, the UKMO 500- mb heights exhibit more ridging downstream of the incipient system but no differences upstream in the vicinity of the upper-level trough. In Fig. 14b, the differences in 850-mb temperatures are quite small in the vicinity of the storm, and larger well upstream, but associated with another system. These initial condition differences are smaller than those for the February case, especially the 850-mb temperatures.

In Fig. 15, the three surface analyses for the AVN, NOGAPS, and UKMO models are shown with available ship and buoy reports. Two incipient lows analyzed by the manual NCEP marine analyses are indicated on the figure. These lows later merge to become the storm. None of the model surface analyses depicts the eastern low at 37.5°N, 178°W that is verified by the ship reports. In subsequent NCEP marine surface analyses, this low is analyzed to weaken over time as it joins the western low at 39°N, 177°E. However, the exact development is difficult to ascertain because of lack of adequate surface reports. Although the forecast sea level pressure errors by all three models at 0000 UTC 12 December 2001 appear to be significant, there is insufficient information to determine whether those errors contributed to the large 48-h forecast errors at 0000 UTC 14 December 2001.

4. Discussion

In this paper, we have demonstrated that large short- term (0 to 48 h) forecast errors of sea level pressure still occur on a regular basis over the northeast Pacific and the west coast. Comparing the NCEP Eta Model forecasts with buoy and CMAN observations over that region for four recent winters, it was found that 48-h sea level pressure errors greater than 10 mb occur 10– 20 times per year, while 24-h errors of that amount typically occur 3–6 times per year. Looking at major wintertime forecast failures over the region (48-h SLP errors greater than two standard deviations from the mean for two or more forecast cycles at two or more observing locations) indicates that 16–21 major forecast failures occur per winter. The vast majority of these problematic forecasts were for surface lows, which possessed average position and mean absolute intensity errors of 453 (260) km and 7.5 (4.1) mb, respectively, for 48-h (24-h) forecasts. These results are consistent with those of Colle and Mass (2001), who found large short- term forecast errors in trough timing by the Eta Model over the eastern Pacific (winter seasons 1997–2000), with timing errors for 3–18-h forecasts ranging as high as 9–12 h.

There appears to be a modest improving trend in the Eta forecasts over the four recent winters. When the two more active winters are compared, 2001/02 and 1999/ 2000, there were fewer large errors in the 2001/02 season compared to the 1999/2000, even though the 2001/ 02 season exhibited greater synoptic activity over the northeastern Pacific and west coast. Perhaps this skill increase reflects improved availability and/or use of data assets in the EDAS or the change in model resolution in September 2000. Even with such improved performance, Eta analyses and subsequent forecasts tend to be inferior to those of the NCEP Global Forecast System (GFS, formerly the AVN) model.4

An important finding of this work is that over the northeast Pacific there are often large differences in the initializations and subsequent forecasts of the major modeling centers. Such differences were evident in the 7 February 2002 case and have been observed by the authors to occur quite frequently even today (early 2003). This variability in initializations, and the subsequent impact on forecast skill, is highlighted by the results of the UW ensemble system, in which the MM5 is driven by different operational initializations and boundary conditions (Grimit and Mass 2002). Specifically, the UW ensembles evince a great deal of variation in the initializations, with some forecasting systems (NCEP GFS and UKMO Unified Model) regularly appearing to be superior. The definitive source(s) of these often large differences are not clear, although variations in assimilation approaches [ranging from optimal interpolation (OI) to 3DVAR], data assets (substantial differences in the use of satellite information), quality control algorithms (particularly criteria determining the acceptance or rejection of observational data), and the realism of the modeling systems (dependent on numerics, resolution, and physical parameterizations) could all provide part of the answer.

With the relatively frequent occurrence of both large short-term model errors and substantial differences in initializations over the eastern Pacific, forecasters not only require better forecasting systems (models, data assimilation) but also improved tools and data for evaluating initialization and short-term forecast quality. The need for such tools and data is highlighted by faulty short-term forecasts in which the public are not only provided a poor prediction, but no foreknowledge that the forecast was unreliable (e.g., 7 February 2002, 3 March 1999). To address this problem, West Coast forecasters require software that would objectively compare model initializations and forecasts against all observational assets. A project at the University of Washington has built a prototype of such a system that compares the output of several models with a wide range of observations over the eastern Pacific (http://www.atmos.washington.edu/~bnewkirk).

Another problem is that east Pacific observing assets are inadequate for evaluating forecast accuracy of landfalling systems. For example, the Pacific Northwest has the worst coastal radar coverage in the lower 48 states (Westrick et al. 1999). The placement of Weather Surveillance Radar-1988 Doppler (WSR-88D) units (or equivalent) on the coast would provide a view of incoming systems for several hundred kilometers offshore, allowing forecasters 6–12 h of warning of major forecast problems prior to landfall and would act as a last line of defense for poorly forecast events. Additional coastal buoys within a few hundred kilometers of the coast would also be highly valuable.

An important issue regarding large forecast errors over the northeast Pacific (and elsewhere) is whether short-term prediction failures are associated with large initial errors that grow or small errors in regions of great sensitivity that amplify rapidly. The differences in the initial state were large for the 7 February 2002 event, and the 24-h forecasts verifying at time of landfall were poor. The fact that even very short-term forecasts (6 to 12 h) were seriously in error for this case indicated a major failure of operational data assimilation systems to correctly depict the incipient storm. In contrast, the initial condition errors for the December event appeared to be small, except possibly at the surface, and the shorter-term forecasts (less than or equal to 24 h) for this event were relatively accurate. Although we have incomplete evidence, it is possible that in the December case, rapid growth of small errors contributed to the large 48-h forecast errors. However, note that for both events, there were very few ship and buoy observations in the vicinity of the developing storms, inhibiting the ability to accurately determine the initial condition error.

Another major question is the relative importance of initialization error versus model error (deficiencies in the numerical methodology or physics parameterizations). The fact that ensemble runs of the February storm using the same model but different initial conditions closely parallel the parent operational models in which they were nested suggests the dominance of initial condition variability over model physics uncertainty in the February case.

Paradoxically, the continued existence of large initialization errors and their association with poor short- term forecasts suggests the potential for substantially improved forecasts. As new observing systems, better quality control procedures, and improved data assimilation approaches improve initialization quality, such blatantly incorrect initial states should lessen in frequency, with improved forecasts as a consequence. Such events are the “low-hanging fruit” for the upcoming THORPEX Project, which will develop new observational and data assimilation approaches for addressing initialization deficiencies over the world's oceans. Improvement of forecasts in which small initialization errors occur in regions of great sensitivity, thus resulting in rapid error growth, is far more difficult.

5. Summary

This paper documents the magnitude and frequency of large forecast errors of sea level pressure over the northeastern Pacific and the U.S. West Coast during four recent winter seasons (October 1999–March 2003). In addition, two poorly forecast storms that had substantial impact over the region were examined. The findings can be summarized as follows:

  • Large short-term forecast errors occur several times each winter over the northeast Pacific in the Eta Model. Specifically, Eta Model 48-h sea level pressure forecast errors greater than 10 mb and Eta Model 24- h sea level pressure forecast errors greater than 6 mb both occur 10–20 times each winter season at northeastern Pacific buoy and coastal locations. Eta Model 48-h sea level pressure forecast errors greater than 15 mb and Eta Model 24-h sea level pressure forecast errors greater than 10 mb occur 3–6 times per winter season over the same area.
  • Eta Model error variability is modulated by synoptic activity, with larger numbers of major forecast errors in more active years.
  • Considering interannual variability, there appears to be an improving trend for Eta forecasts, with fewer major forecast failures during the active 2001/02 season, compared to the 1999/2000 season, which had less storm activity.
  • Large initialization errors over the Pacific still occur in all operational forecast systems and are major contributors to the failures of short-term (0–48 h) forecasts.
  • Large sea level pressure forecast errors at northeast Pacific buoy and coastal stations are generally associated with large position errors of surface low pressure centers.
  • Large initial condition errors played a major role in the large forecast errors of the 7–8 February 2002 case. The operational model forecasts with the largest 12–24-h forecast errors for the 7–8 February 2002 storm had initial conditions that failed to depict the correct location or sharpness of the incipient system.
  • A significant number of large short-term forecasting failures documented in this paper can be traced to poor initializations over the Pacific. New observing assets and better data assimilation approaches, which are important components of the upcoming THORPEX experiment, may well be able to improve these significant forecast failures.

Acknowledgments

This work was supported by the DoD Multidisciplinary University Research Initiative (MURI) program, administered by the Office of Naval Research under Grant N00014-01-10745, and by Office of Naval Research Grant N00014-98-1-0193. The authors also extend their thanks to Drs. Greg Hakim and Rolf Langland for their comments on this paper and for many fruitful discussions on predictability issues. We also acknowledge Mr. Brian Ancell for providing MM5 ensemble runs for the February case, Mr. David Ovens for developing and maintaining the MM5 and Eta Model verification systems, and Dr. Jeff Yin for providing software used in creating Fig. 5. NCEP reanalysis data was provided by the NOAA–CIRES Climate Diagnostics Center, Boulder, Colorado, from their Web site (http:// www.cdc.noaa.gov).

REFERENCES

  • Anthes, R. A., , Kuo Y-H. , , and Gyakum J. R. , 1983: Numerical simulations of a case of explosive cyclogenesis. Mon. Wea. Rev., 111 , 11741188.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Black, T. L., 1994: The new NMC mesoscale Eta Model: Description and forecast examples. Wea. Forecasting, 9 , 265278.

  • Bosart, L. F., 1981: The Presidents' Day snowstorm of 18–19 February 1979: A subsynoptic-scale event. Mon. Wea. Rev., 109 , 15421566.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Bosart, L. F., , and Lin S. C. , 1984: A diagnostic analysis of the Presidents' Day storm of February 1979. Mon. Wea. Rev., 112 , 21482177.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Buizza, R., , and Chessa P. , 2002: Prediction of the U.S. storm of 24– 26 January 2000 with the ECMWF Ensemble Prediction System. Mon. Wea. Rev., 130 , 15311551.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Colle, B. A., , and Mass C. F. , 2001: Evaluation of the timing and strength of MM5 and Eta surface trough passages over the eastern Pacific. Wea. Forecasting, 16 , 553572.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Colle, B. A., , Westrick K. J. , , and Mass C. F. , 1999: Evaluation of the MM5 and Eta-10 precipitation forecasts over the Pacific Northwest during the cool season. Wea. Forecasting, 14 , 137154.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Duchon, C. E., 1979: Lanczos filtering in one and two dimensions. J. Appl. Meteor., 18 , 10161022.

  • Gelaro, R., , Reynolds C. A. , , Langland R. H. , , and Rohaly G. D. , 2000:: A predictability study using geostationary satellite wind observations during NORPEX. Mon. Wea. Rev., 128 , 37893807.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grimit, E. P., , and Mass C. F. , 2002: Initial results of a mesoscale short-range ensemble forecasting system over the Pacific Northwest. Wea. Forecasting, 17 , 192205.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Grumm, R. H., , and Siebers A. L. , 1989: Systematic surface cyclone errors in NMC's nested grid model November 1988–January 1989. Wea. Forecasting, 4 , 246252.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Gyakum, J. R., 1983a: On the evolution of the QE II storm. I: synoptic aspects. Mon. Wea. Rev., 111 , 11371155.

  • Gyakum, J. R., 1983b: On the evolution of the QE II storm. II: Dynamic and thermodynamic structure. Mon. Wea. Rev., 111 , 11561173.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Klinker, E., , Rabier F. , , and Gelaro R. , 1998: Estimation of key analysis errors using the adjoint technique. Quart. J. Roy. Meteor. Soc., 124 , 19091933.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Kuo, Y-H., , and Reed R. J. , 1988: Numerical simulation of an explosively deepening cyclone in the eastern Pacific. Mon. Wea. Rev., 116 , 20812105.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Langland, R. H., and Coauthors, 1999: The North Pacific Experiment (NORPEX-98): Targeted observations for improved North American weather forecasts. Bull. Amer. Meteor. Soc., 80 , 13631384.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Langland, R. H., , Shapiro M. A. , , and Gelaro R. , 2002: Initial condition sensitivity and error growth in forecasts of the 25 January 2000 East Coast snowstorm. Mon. Wea. Rev., 130 , 957974.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mass, C. F., , Ovens D. , , Westrick K. , , and Colle B. A. , 2002: Does increasing horizontal resolution produce better forecasts? The results of two years of real-time numerical weather prediction in the Pacific Northwest. Bull. Amer. Meteor. Soc., 83 , 407430.

    • Search Google Scholar
    • Export Citation
  • Miguez-Macho, G., , and Paegle J. , 2000: Sensitivity of a global forecast model to initialization with reanalysis datasets. Mon. Wea. Rev., 128 , 38793889.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Mullen, S. L., 1994: An estimate of systematic error and uncertainty in surface cyclone analysis over the North Pacific Ocean: Some forecasting implications. Wea. Forecasting, 9 , 221227.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • NCDC, 2001: Storm Data. Vol. 43, No. 12, 106 pp. [Available from W. Angel, Ed., National Oceanographic and Atmospheric Administration, National Climatic Data Center, Asheville, NC 28801-2733.].

    • Search Google Scholar
    • Export Citation
  • NCDC, 2002: Storm Data. Vol. 44, No. 2, 90 pp. [Available from W. Angel, Ed., National Oceanographic and Atmospheric Administration, National Climatic Data Center, Asheville, NC 28801- 2733.].

    • Search Google Scholar
    • Export Citation
  • Nelsen, J. A., 1999: The Eta data assimilation system. WR Tech. Attachment 99-14, 6 pp. [Available from National Weather Service Western Region, P.O. Box 11188, Salt Lake City, UT 84147.].

    • Search Google Scholar
    • Export Citation
  • Oravec, R. J., , and Grumm R. H. , 1993: The prediction of rapidly deepening cyclones by NMC's nested grid model: Winter 1989– autumn 1991. Wea. Forecasting, 8 , 248270.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rabier, F., , Klinker E. , , Courtier P. , , and Hollingsworth A. , 1996: Sensitivity of forecast errors to initial conditions. Quart. J. Roy. Meteor. Soc., 122 , 121150.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Reed, R. J., , and Albright M. D. , 1986: A case study of explosive cyclogenesis in the eastern Pacific. Mon. Wea. Rev., 114 , 22972319.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Black T. , , Collins W. , , Manikin G. , , Mesinger F. , , Parrish D. , , and DiMego G. , 2000: Changes to the NCEP Meso Eta Analysis and Forecast System: Assimilation of satellite radiances and increase in resolution. NWS Technical Procedures Bulletin 473, 10 pp. [Available online at http://wwwt.emc.ncep.noaa.gov/ mmb/mmbpll/eta22tpb/; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Black T. , , Ferrier B. , , Lin Y. , , Parrish D. , , and DiMego G. , 2001a: Changes to the NCEP Meso Eta Analysis and Forecast System: Increase in resolution, new cloud microphysics, modified precipitation assimilation, modified 3DVAR analysis. NWS Technical Procedures Bulletin 488, 15 pp. [Available online at www.emc.ncep.noaa.gov/mmb/mmbpll/eta12tpb; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

    • Search Google Scholar
    • Export Citation
  • Rogers, E., , Ek M. , , Lin Y. , , Mitchell K. , , Parrish D. , , and DiMego G. , 2001b:: Changes to the NCEP Meso Eta Analysis and Forecast System: Assimilation of observed precipitation, upgrades to land-surface physics, modified 3DVAR analysis. NWS Technical Procedures Bulletin 479, 20 pp. [Available online at www.emc.ncep.noaa.gov/mmb/mmbpll/spring2001/tpb/; also available from National Weather Service, Office of Meteorology, 1325 East-West Highway, Silver Spring, MD 20910.].

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

  • Shapiro, M., , and Thorpe A. , cited 2002: Program overview: The observing-system research and predictability experiment, THORPEX. [Available online at www.mmm.ncar.edu/uswrp/thorpex/thorpex_plan13.pdf.].

    • Search Google Scholar
    • Export Citation
  • Smith, B. B., , and Mullen S. L. , 1993: An evaluation of sea level cyclone forecasts produced by NMC's nested-grid model and global spectral model. Wea. Forecasting, 8 , 3756.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Steenburgh, W. J., , and Mass C. F. , 1996: Interaction of an intense extratropical cyclone with coastal orography. Mon. Wea. Rev., 124 , 13291352.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Szunyogh, I., , Kalnay E. , , Morss R. E. , , Majumdar S. J. , , Etherton B. J. , , and Bishop C. H. , 2000: The effect of targeted dropsonde observations during the 1999 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 128 , 35203527.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Szunyogh, I., , Toth Z. , , Zimin A. V. , , Majumdar S. J. , , and Persson A. , 2002: Propagation of the effect of targeted observations: The 2000 Winter Storm Reconnaissance Program. Mon. Wea. Rev., 130 , 11441165.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., 1986: The possible influence of upstream upper- level baroclinic processes on the development of the QEII storm. Mon. Wea. Rev., 114 , 10191027.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., , Kocin P. J. , , Petersen R. A. , , Wash C. H. , , and Brill K. F. , 1984:: The Presidents' Day cyclone of 18–19 February 1979: Synoptic overview and analysis of the subtropical jet streak influencing the pre-cyclogenetic period. Mon. Wea. Rev., 112 , 3155.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Uccellini, L. W., , Keyser D. , , Brill K. F. , , and Wash C. H. , 1985: Presidents' Day cyclone of 18–19 February 1979: Influence of upstream trough amplification and associated tropopause folding on rapid cyclogenesis. Mon. Wea. Rev., 113 , 962988.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Westrick, K., , Mass C. , , and Colle B. , 1999: The limitations of the WSR-88D radar network for quantative precipitation measurement over the coastal western United States. Bull. Amer. Meteor. Soc., 80 , 22892298.

    • Crossref
    • Search Google Scholar
    • Export Citation
  • Zhang, F., , Snyder C. , , and Rotunno R. , 2002: Mesoscale predictability of the “surprise” snowstorm of 24–25 January 2000. Mon. Wea. Rev., 130 , 16171632.

    • Crossref
    • Search Google Scholar
    • Export Citation

Fig. 1.
Fig. 1.

NCEP Eta Model 24-h (solid) and 48-h (dashed) forecast of SLP (mb) verifying at 0000 UTC 3 Mar 1999. The corresponding GOES-10 infrared satellite image is also shown. The analyzed surface low location and estimated central pressure from subjective surface analyses are indicated by the larger “L.”

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 2.
Fig. 2.

Map of buoy (e.g., 46001) and coastal (e.g., DESW1) station locations used in the study

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 3.
Fig. 3.

Time series of SLP errors defined as the forecast SLP minus the observed SLP (mb) at the station. The average error at each station for each forecast time is plotted as a solid line and plus/minus 2 times the std dev is shown as dashed lines. Arrows and numbers 1, 2, and 3 refer to specific events discussed in the text. (a) The 24-h forecast errors at CMAN station Tatoosh Island, WA (TTIW1), (b) 24-h forecast errors at CMAN station Cape Arago, OR (CARO3), (c) 48-h forecast errors at Tatoosh Island, and (d) 48-h forecast errors at Cape Arago

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 4.
Fig. 4.

Histograms of 48-h forecast SLP error (mb) at Tatoosh Island, for the four winter seasons. The average error and std dev are provided for each season

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 5.
Fig. 5.

Mean 500-mb heights and rms calculated from time-filtered (2–10 days) 500-mb heights over the North Pacific and coastal regions for the (a) 1999/2000 winter season, (b) 2000/01 winter season, (c) 2001/02 winter season, and (d) 2002/03 winter season

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 6.
Fig. 6.

Histograms of cyclone central pressure and position errors for the events listed in Table 4. The average errors and mean absolute errors are also indicated: (a) 24-h and (b) 48-h forecasts

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 7.
Fig. 7.

The 48-h forecasts of SLP (solid) and 00-h AVN analysis of SLP (dashed) overlaid on the infrared satellite image valid 0000 UTC 8 Feb 2002 by the (a) AVN, (b) Eta, (c) UKMO,  and (d) NOGAPS models

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 8.
Fig. 8.

As in Fig. 7, except for 24-h forecasts of SLP valid 0000 UTC 8 Feb

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 9.
Fig. 9.

Initial condition differences between the UKMO and the Eta analyses at 0000 UTC 7 Feb 2002. (a) The 500-mb heights from the UKMO (solid) and Eta (dashed) and the difference (ETA − UKMO; shaded). The differences are shaded every 10 m starting at 30 m. In addition, the trough lines from the UKMO heights and Eta heights are shown with solid and dashed black lines, respectively. (b) The 850-mb temperature (K) from the UKMO (solid) and Eta (dashed), with the positive differences (UKMO − Eta) in shades of gray every 0.5°C and negative differences contoured in gray every 0.5°C

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 10.
Fig. 10.

Initial SLP fields and available surface reports for the (a) AVN, (b) Eta, (c) UKMO, and (d) NOGAPS model analyses for 0000 UTC 7 Feb 2002. The corresponding GOES-W infrared satellite image is also shown

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 11.
Fig. 11.

The 24-h forecasts of SLP (solid) and the AVN analysis of SLP (dashed) overlaid on the infrared satellite image valid at 0000 UTC 8 Feb 2002 by the (a) MM5-AVN, (b) MM5- Eta, (c) MM5-UKMO, and (d) MM5-NOGAPS

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 12.
Fig. 12.

The 48-h forecasts of SLP (solid) and 00-h AVN SLP analysis (dashed) overlaid on the infrared satellite image valid at 0000 UTC 14 Dec 2001 by the (a) AVN, (b) Eta, (c) UKMO, and (d) NOGAPS models

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 13.
Fig. 13.

(a) Forecast error of cyclone central pressure (mb) and (b) cyclone position (km) as a function of forecast lead time verifying at 0000 UTC 14 Dec 2001 from the AVN, CMC, Eta, NOGAPS, and UKMO models. The verifying position and central pressure is obtained from subjective analysis

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 14.
Fig. 14.

Initial condition differences for the 48-h forecasts initialized at 0000 UTC 12 Dec. (a) The 500-mb heights (m) produced by the UKMO (solid) and the NOGAPS (dashed) models and the difference (UKMO − NOGAPS; shaded). The differences are shaded in gray every 10 m starting at 30 m. (b) The 850-mb temperature (K) from the UKMO (solid) and NOGAPS (dashed) models and the positive differences (UKMO − NOGAPS) in shades of gray every 0.5°C and negative differences contoured in gray every 0.5°C

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Fig. 15.
Fig. 15.

SLP analyses and ship and buoy surface reports at 0000 UTC 12 Dec produced by the (a) AVN, (b) NOGAPS, and (c) UKMO models. Location of incipient surface lows obtained from the NCEP marine surface analyses are indicated by the two larger “L”s at 37.5°N, 178°W and 39°N, 177°E

Citation: Weather and Forecasting 19, 2; 10.1175/1520-0434(2004)019<0338:MNFFOT>2.0.CO;2

Table 1.

Number of Eta Model forecasts where the SLP errors are greater than 2 times the std dev (3 times the std dev in parentheses) for 24-h forecast errors. The stations are listed geographically from north to south. See Fig. 2 for locations of stations

Table 1.
Table 2.

Number of Eta Model forecasts where the SLP errors are greater than 2 times the std dev (3 times the std dev in parentheses) for 48-h forecast errors. The stations are listed geographically from north to south. See Fig. 2 for locations of stations

Table 2.
Table 3.

Interannual variability of the large Eta Model forecast errors presented in Tables 1 and 2. Average number of 24- and 48- h errors greater than 2 and 3 std dev (SD) year−1 for those observing sites in Tables 1 and 2 that had complete records (46027, CARO3, 46005, DESW1, TTIW1, 46001, and 46184)

Table 3.
Table 4.

Total number of events with large Eta Model 48-h forecast errors at northeast Pacific buoys and coastal sites for each winter season. Large events are defined as times when SLP errors are greater than 2*SD for two or more consecutive forecasts at two or more adjacent stations. Also shown are the number of these events asso ciated with surface lows, troughs, and high pressure systems. The event was considered a “low” when the observing stations with errors were in the vicinity of a low, or closed contours of SLP, and the event was considered a trough if the observing stations were in the vicinity of a frontal trough and far away from the parent surface low

Table 4.
Table 5.

A list of the operational models and their resolutions and objective analysis schemes used at the time of the Dec 2001 and Feb 2002 storms

Table 5.

1

Data from stations 46001 and 46184 were not included for identifying events since they are located far from the other buoys included in the study.

2

These sites include buoys 46027, 46005, 46001, and 46184, and coastal stations Cape Arago (CARO3), Destruction Island (DESW1), and Tatoosh Island (TTIW1).

3

The Lanczos filter used for the rms calculations is given by Duchon (1979).

4

In a preliminary study by the authors, the 00-, 24-, and 48-h forecasts of SLP by the GFS model had smaller mean absolute errors and standard deviations than the Eta Model at 11 of the 16 stations for the 2002/03 winter season.

Save