Close Window

Pennsylvania State University (Robert Hart, COMET graduate student fellow): "Forecasting studies using hourly model-generated soundings"
Final Report

I. Introduction

This report is the final report which summarizes the research and findings of the cooperative project between Robert Hart, Dr. Gregory S. Forbes, and the National Weather Service in State College Pennsylvania, on the use of hourly model-generated profiles to aid forecasting of mesoscale atmospheric phenomena. A summary of the original goals of the project are presented, along with any modifications to those goals that were necessary along the course of the research. Secondly, a summary of the progress to date of the research is presented, including the lessons learned during the research, a summary of interactions with the National Weather Service, a description of the tools developed, and an explanation of any problems encountered. Finally, a brief description of any work remaining for this project is given along with a concluding summary.

II. Summary of Original Goals and any modifications

A. The primary goal of this research was to examine how high-resolution forecast vertical profiles of the atmosphere could be used to increase the forecast accuracy of mesoscale phenomena. The vertical profiles were made available to us through
B. NCEP's anonymous-ftp server for each of NCEP's operational models: NGM, Eta, and Mesoscale Eta. Each of the specific goals are described below:
C. Decode, and generate user-friendly graphical forecast output from the vertical profiles.
D. Make the graphical output available to forecasters at the National Weather Service and to forecasting classes and other groups
E. Through cooperation with the National Weather Service, evaluate the graphical products for their utility in the forecasting process.
F. Perform both subjective and objective statistical analyses on the high-resolution data sets to determine how these products can be most accurately used.
G. Generate experimental products to aid in forecasting phenomena that are typically most difficult to forecast.
H. Present and document findings at Weather and Forecasting Conferences, in Hart's M.S. Thesis, and in journal articles.<

No modifications to the original goals were necessary.

III. Summary of Progress to Date

A. Scientific lessons learned

During the past two years of research for this COMET Fellowship, numerous scientific lessons were learned. These lessons are outlined and described in detail below, with further explanation and examples given in Hart's M.S. thesis.


1. The data set used, hourly model-generated soundings, is a highly valuable source of numerical guidance in operational forecasting. Most importantly, the datasheet and corresponding graphics enhance the forecasters' ability to visualize the atmosphere and make forecasts without overwhelming them. Further, in a survey conducted on students in the Department of Meteorology; the model soundings were rated second most valuable (desirable) model data set.

2. The model soundings, while valuable, need to be used in conjunction with and in the context of the lower-resolution two-dimensional grids. Only when the two types are used together is the utility of the model soundings greatest. In fact, using the model soundings alone to forecast can be extremely detrimental since forecasters are less able to judge model timing errors.

3. The high resolution time-height cross sections derived from the model soundings are an outstanding method for visualizing and teaching the atmosphere. This conclusion was drawn not only from Hart's daily usage of the graphics, but also through correspondence with other high school and university teachers and professors who had great praise for the graphics' utility. In fact, several teachers have incorporated the products into their basic and advanced courses on meteorology and atmospheric phenomena.

4. Users of the model soundings need to know ahead of time the exact location of the model-sounding site. Since the model sounding represents the nearest grid point (and not an interpolation), significant displacements in the horizontal and vertical are possible between the model sounding and the forecast location. This influence is especially significant if the model sounding site is located over water while the forecast site is over land, or vice-versa. The surface type (water or land) has a dramatic impact on the forecasts of precipitation type, temperature, and frost/freeze conditions.


1. The Eta and Mesoscale Eta models are, in general, more accurate at predicting low-level temperature, moisture, and wind speeds than the predecessor (Nested Grid Model, NGM).

2. The precipitation type algorithm included as part of the Eta and Mesoscale Eta model output provides for, on average, more accurate predictions of precipitation type and changeover than was available from the NGM Model-Output-Statistics (MOS). In particular, precipitation type forecasts during extreme or anomalous events were far more accurately predicted using the Eta and Mesoscale Eta models than using the NGM MOS. On average, however, the precipitation type scheme tends to over predict snow at lower elevations and over predict sleet and freezing rain at higher elevations (> 1000 feet).

3. Despite the increased accuracy, there remain significant biases associated with the Eta and Mesoscale Eta Models. During clear, calm nights (especially during the winter), these models tend to overcool the boundary layer, leading to 2-meter (surface) temperatures that are several degrees too cold. Such biases have dramatic impact on forecasts of precipitation type, low temperatures, and the experimental frost/freeze forecasts. The bias produces excessive false-alarm rates for freezing rain. Critical success indices for freezing rain and sleet remain very low.

4. During the first 18 months of research, a parameterization error in the Eta and Mesoscale Eta models produced an excess of net radiation at the surface. This produced, on average, higher afternoon temperatures than observed a deeper boundary layer, and lower surface dew points than observed. All these biases had a strong influence on predicting convective initiation, potential convective intensity (CAPE), and afternoon high temperatures. Fortunately, this bias was fixed in early 1997. However, researchers using archived past data sets prior to this date should be aware of this bias.

5. During times of easterly flow, the model often veers the low-level wind too greatly, often producing a forecast for too strong a southerly or westerly component. This has a significant impact on low-level moisture, cloudiness forecasts, and environmental helicity forecasts, all of which greatly impact forecasts of convective potential and storm type/severity.

6. Model-derived Richardson Number forecasts can be used to predict pilot-reported turbulence with reasonable success, especially in the boundary layer. It was found that 66% of the time a pilot reported turbulence, the model forecasted a Richardson number less than the empirical threshold of 0.25. Further, nearly 90% of the time a pilot reported turbulence the forecast value was one or less. As described in detail in the M.S. Thesis, the depth of the forecast potentially turbulent layer (Richardson number < 1) was found to have a strong influence on the occurrence of turbulence. Below a certain threshold (typically, 50-75mb) depth, turbulence was not reported regardless of how unstable the layer was.

7. The experimental wind gust forecast product performs well in isolating windows of potentially strong to damaging winds. The product is able to distinguish the threat to elevated regions from the threat to those regions in protected valleys. The greatest sensitivity and cause for error is in the accuracy of the model forecasts of low-level stability. If the model inaccurately forecasts the low-level stability, the experimental wind gust forecast is greatly influenced since the boundary layer wind speeds are affected. Thus, forecasters must constantly compare model to reality to determine if the atmosphere is actually more or less conducive to damaging winds than the guidance indicates.

8. The experimental mesoscale precipitation banding product performs well to isolate areas of the forecast region and forecast times that are most conducive to banding potential (through either frontogenesis or conditional symmetric instability); however, there remains significant work to be done in actually forecasting the exact location, duration, and intensity of these bands.

9. The convective parameterization in the Eta and Mesoscale Eta models, the Betts-Miller parameterization, is an undesirable parameterization for mid-latitude convective systems. The scheme is too slow to respond to convective instability, often producing rainfall rates that are an order of magnitude less than observed and several hours late. Further, the scheme cannot respond to elevated instability above the boundary layer. During such cases, the scheme responds too strongly to the low-level unstable air in the warm sector, which then "robs" moisture from the areas of elevated instability, typically to the north.

[Section B: Summary of interactions with NWS omitted here]

C. Description of tools developed

During the course of this research, several tools were developed which have been utilized and extensively evaluated. Each of these tools are described in detail below:

Development of software to automate generation of high-resolution graphical output:

During the first year of this research, a robust and dependable set of routines were developed which, for each of four to six model runs, retrieves the raw hourly profiles from NCEP, decodes them into ASCII tabular format, and generates high-resolution color graphical output for use by the forecasters. Further, the tabular model profiles are compressed and archived. This archive currently extends for 12-18 months, depending on model run.

Development of the first SkewT/LogP routine for the GRADS software package:

The GRADS (Gridding Analysis and Display System) software visualization package was used for graphical representation of the hourly model forecasts. At the time the research was begun, there was no method for displaying SkewT/LogP plots within GRADS. Since SkewT/LogP plots were a highly demanded graphical format requested by both students and the National Weather Service, a GRADS function was written by Hart to generate such plots. This routine is full-featured and its use has expanded considerably to include other universities.

Generation of a World-Wide-Web forms-based interface to the graphical output:

Initially, one of the fundamental problems encountered during this research was how to make available to forecasters the graphical model output. It became clear that the World-Wide-Web would provide an excellent method by which both students and forecasters alike could view the model output. Therefore, once (i) through (ii) above were completed, a web-based forms interface to the graphical data (in GIF format) was created. This interface is user friendly and allows forecasters to view any set of model output from the previous 24 hours. This allows forecasters to perform site-to-site comparisons as well as run-to-run comparisons to view model trends. Usage of the page averages 100 "hits" per day, with usage exceeding 200-300 per day during times of active weather. The URL for this page is

Creation of several experimental forecast product:

One of the original goals of this research was to generate experimental forecasting products from the high-resolution model data. These products would take advantage of the fine resolution (both temporal and spatial) of the profiles. Several experimental products were created during the two years, and these are described briefly below:

1. Fog Forecast

Using the 2-meter parameterized forecast relative humidity, along with the lowest-model-level relative humidity, experimental forecasts of relative humidity were produced for each model run. Depending on the model type, a relative humidity value of 92-95% or higher was indicative of a forecast visibility of 2 miles or less. Forecasting skill of this scheme ranged between 55 and 65%, with greatest forecast error during the early fall and late spring seasons.

2. Wind Gust Forecast

Using the high-resolution boundary layer forecasts of wind speed, experimental forecasts of surface wind gust speed probability were generated. The forecasts were for probability of occurrence of surface wind gusts in excess of 30, 40, and 50mph. The forecasts give forecasters predicted windows of potential strong to damaging winds, with considerable lead time for issuance of advisories or warnings.

3. Frost/Freeze Forecast

The Eta and Mesoscale Eta hourly profiles provide forecasts of 2-meter temperature and dewpoint, along with skin temperature. Using these forecast values, relatively successful experimental forecasts of frost or freeze conditions and duration are available via the web page. Forecasts for frost were made if the forecast skin temperature fell below the 2-meter dewpoint. Forecasts for freezing conditions were made if the skin temperature and the 2-meter temperature both fell below freezing.

4. Mesoscale Precipitation Banding Forecast

The most ambitious experimental tool was the creation of a forecasting product for mesoscale banding of precipitation. Such forecasts are made available as two-dimensional (x-y) fields of (1) forecast vertically integrated frontogenesis and (2) atmospheric depth that is saturated and symmetrically unstable. Preliminary results indicate that these products give forecasters an excellent view of the part of the local area that is susceptible to mesoscale banding; however, there remains a great deal of work to be done in predicting the actual location and intensity of such mesoscale bands.

5. Turbulence Forecasting

The Eta and Mesoscale Eta models are the first models that can predict Richardson Number to values of less than 0.5. Previous models (LFM, NGM) had insufficient vertical resolution to resolve stability and wind velocity gradients that could produce such low values of Richardson Number.

Detailed descriptions of these experimental tools are presented in Hart's M.S. Thesis.

Generation of a World-Wide-Web page for browsing of high-resolution real-time meteorological fields which synthesize hourly model forecast and surface data.

During the mid 1990s, Forbes created, on the Penn State University Meteorology Computer System, a real-time system for display of convective parameters, such as Moisture Convergence fields, Lifted Index calculations, and surface dewpoint. During the second year of this research, this idea was expanded upon considerably and the graphical output was greatly improved using GRADS. The National Weather Service in State College now routinely uses the hourly surface analyses via the World-Wide-Web, at Available from this web page are hourly plots of Forbes' original page, but also plots of estimated CAPE (convective available potential energy), CIN (Convective inhibition), and surface horizontal frontogenesis. The hourly model soundings are used as the basis for Lifted Index, CAPE, and CIN calculations. Several forecasters at the office very enthusiastically use these plots during the convective season, and support from the Internet community has been strong.