Close Window

"An investigation of a new operational tool for forecasting downslope windstorms": Louisa Bogar Nance, COMET Postdoctoral Fellow (Seattle Forecast Office)

Final Report

Executive Summary

The ability of a two-dimensional numerical model initialized with upstream profiles taken from Eta gridded data to forecast downslope windstorms was investigated by considering high wind events that affected Boulder, CO between December 1993 - April 1996. These test cases indicate this approach shows some promise. Real-time simulations generated for the Boulder configuration between October 1997 - April 1998 revealed the configuration of the model applied in this region has an unacceptably high false warning rate. Changing the location from which the forecast wind speed was obtained and the addition of a simple surface friction parameterization were able to significantly reduce the false warning rate, but the false warning rate was still unacceptably high. Denver soundings corresponding to the time of the false warnings suggest the model's high false warning rate may be due to its inability to account for a cold pool downstream of the mountain ridge. This shortcoming might be remedied by simply revising the manner in which the model is initialized, which is still under investigation.

The model's ability to forecast downslope windstorms was also investigated by considering high wind events that affected Enumclaw, WA; Salt Lake City, UT; Juneau, AK; Anchorage, AK; Fairbanks, AK; and Great Falls, MT between 1993 and 1997. The initialization of the two-dimensional model was revised slightly prior to generating simulations for these regions. Based on the results from these test cases, the model's ability to forecast downslope windstorms that affect Enumclaw, Salt Lake City, and Juneau looks very promising. The results for Anchorage, Fairbanks, and Great Falls were also promising, but these results suggest the model may have some trouble predicting the strength of high wind events in these regions. To further evaluate the model's ability to forecast high wind events in these six regions, the model will be run in real-time on a workstation in these six forecast office through Spring 1999.

In addition to investigating this new approach to forecasting downslope windstorms, I took part in a number of activities geared towards familiarizing myself with the forecast process and promoting interaction between myself and the operational forecasters. To familiarize myself with the forecast process, I observed lead forecasters at the Seattle forecast office during their shifts. To promote interaction between myself and the forecasters at the Seattle office, I attended daily map briefings and various other office activities, and presented brown bag seminars on terrain-induced flows. To promote interaction between myself and forecasters at other offices whose CWAs are affected by downslope windstorms, I organized a workshop on downslope windstorms that was held 14-15 November 1996, in Seattle, and visited each of the forecast offices that will be participating in the real-time evaluation of numerical model. Working at a local forecast office for the past two years definitely enhanced my overall experience.

Final Report

The development of the model code I was originally planning to use for this study was running behind schedule, so Professor Durran and I decided it would be best for me to work with the code I used during my graduate studies. Since this code was setup to run on the Cray Y-mp at NCAR, the first task I undertook after transferring the code to a HP workstation located at the Seattle forecast office was the removal of Cray-dependencies. These modifications entailed locating HP functions that could be substituted for Cray functions, and setting up a totally new output routine. With the help of Scott Jacobs at NCEP, I was able to put together a new output routine that writes the model fields directly to a GEMPAK file. After making these modifications, I tested the new version of the code by running a few simple test cases using a Witch of Agnesi mountain profile and background flows with constant stability and wind speed. Once all the minor bugs were fixed, I developed a new input routine that allows the model to be initialized with profiles retrieved from Eta forecast grids. With this new input interface, the user initiates a simulation by simply specifying the YYMMDDHH of the model initialization, the forecast hour, and the grid point to be used to initialize the model. A script then takes care of the rest. This setup was tested by generating simulations for a Witch of Agnesi mountain profile and Eta forecast grids corresponding to a high wind event that affected Boulder in April 1996.

During the model development stage, I was also busy with a number of other activities. I made inquiries about high-resolution terrain elevation data sets. It was fairly easy to locate data sets for the lower 48 states, but I ran into some difficulties for the Alaska region, a problem that was eventually solved during my second year. Another major task I addressed during this time period, with help from Eric Thaler (SOO WSFO Denver), was compiling a list of high wind events that affected the Boulder area during the time period December 1993 - April 1996. Once this list was compiled, I submitted a request to have the Eta grids of interest extracted from the archive maintained by the Department of Atmospheric Sciences at the University of Washington. In addition, I spent some time making inquiries about other data archives that might prove useful to this study,.

In June 1996, I attended the first week of the Mesoscale Meteorology Course for University Faculty in Norman, Oklahoma, which was offered by COMET and the University of Oklahoma, and the COMET Symposium / Workshop on Mesoscale Numerical Modeling in Fairbanks, Alaska. I presented an invited talk at the COMET workshop outlining my research project, along with some very preliminary results. Although I was slightly under-the-weather during my trip to Alaska, I felt it was well worth my time since the workshop gave me a chance to heighten the awareness of my research project within the National Weather Service, and, in turn, heighten my awareness of other research projects related to improving operational forecasts.

Upon returning from my excursions the first half of June, I started work on the Boulder windstorm events. Although the model was basically ready to go, I still needed a representative terrain profile to input into the model. After looking at the high resolution terrain elevation data I had obtained for Colorado, and some discussions with Brad Colman and Dale Durran, I decided to use a representative terrain profile based on the maximum terrain elevation along a line parallel to the major mountain ridge axis in a longitudinal band centered on Boulder. The resulting silhouette was smoothed by applying a 9-point box-car average three times, interpolated to the model grid using a cubic spline, and finally reduced by a constant such that the minimum terrain elevation was zero. Initial test simulations were generated using this terrain profile, the 12 and 18 hr forecasts leading the high wind event by three to nine hours, the grid point at 40.35N / 108.30W, and the westerly wind component. While a number of these simulations produced wind speeds in the lee of the mountain ridge that agreed fairly well with the observed wind speeds, a number of simulations produced leeside wind speeds that were either too large or too small. Sensitivity tests in which the projection angle for the cross-barrier component of the large-scale flow, the terrain profile, and the time of the forecast were varied did not improve the agreement between model and observed wind speeds. On the other hand, high wind watches/warnings based on the model output would have captured a larger percentage of these high wind events than the current forecast technique, which, in turn, suggests this approach shows some promise.

To test the question of how well the model performs during both nonwindstorm and windstorm events, the model was set up to run in real time using the current 12 and 18 hr Eta forecasts (four runs daily) beginning 8 October 1996. Due to the almost simultaneous decommissioning of the Colorado mesonet, another source of verification data was needed. A script was set up to retrieve wind speed and direction observations from the NCAR Mesa and Foothills Laboratory weather stations at 15-minute intervals. The performance of the numerical model was then evaluated by comparing the model wind speeds with the maximum magnitude of the observed peak gust for wind directions between 225 and 315 degrees during the time period three to nine hours after the forecast time used to initialize the model. The trends in the model wind speeds between October 8 and November 14 (timing of downslope windstorm workshop) showed promise, but the model wind speeds were still either too large (false alarms in some cases) or too small on a number of occasions.

During the month of October and the first two weeks of November, I was busy juggling research and planning the downslope windstorm workshop that was held 14-15 November 1996, in Seattle, Washington. This workshop provided a nice mixture of short talks by operational forecasters, more extensive talks by two invited speakers, and a number of presentations related to my research project. The presentations related to my project were designed to familiarize the forecasters with the approach I am pursuing toward improving downslope windstorm forecasts, and outline what information I would need from them to test my approach in their CWA. We had a total of 38 attendees with representatives from each region within the National Weather Service.

Following the workshop, I put together a script to automatically generate postscript files containing contour plots of the fields generated by the real-time simulations for the Boulder setup, which continued to run on my workstation four times daily through April. These files were placed in an anonymous ftp directory, which allowed Eric Thaler to retrieve plots of the output on a daily basis. Unfortunately, the feedback from Eric on these real-time runs was rather limited since he was busy juggling work and school.

Another task I undertook following the workshop was putting together information packets on the workshop to be sent to those NWS offices who had expressed interest in the workshop but had been unable to send a representative. These packets contained copies of the overheads from the talks presented by myself and the invited speakers, a video tape of the first day of the workshop (audio is unfortunately missing from the tapes for second day), and a short write-up about the workshop.

Potential problems with the NCAR wind sensor instrumentation were brought to my attention during the downslope windstorm workshop by Dr. John Brown and Eric Thaler. Dr. Brown was kind enough to provide me with wind speed observations from his weather log for the time period 6 October 1996 to 31 December 1996. Comparing these observations with the data from the NCAR Mesa and Foothills Laboratory weather stations, I found the agreement between the Foothills Laboratory wind sensor readings and the entries in Dr. Brown's logbook were much better than originally anticipated. In fact, the agreement was good enough that I decided to use the Foothills data to verify the model forecasts for the entire time period the model was run four times daily (Oct 1996 - April 1997).

Two high wind events and several moderate wind events affected Boulder during the time period between 8 October and 14 November. Unfortunately, due to problems at NCEP, Eta grids were not available for one of the cycles corresponding to the timing of one of the high wind events. The model forecasts based on the available grids were too weak for both high wind events and some of the moderate events. Since shallow inversion layers can play an important role in the generation of strong lee-slope winds, it was suggested at the workshop that the coarse vertical resolution of the Eta grids (50 mb) and the smoothing that takes place when the data is interpolated to the distributed grids may be responsible for the model under forecasting these events. Following Brad's advice, I had started archiving a number of the available Eta bufr soundings (full resolution soundings) that seemed to be ideally located for initializing the model in October. After the workshop, I attempted to address this vertical resolution question by initializing the model with the appropriate bufr soundings. Initializing the model with bufr soundings required some fairly simple modifications to my input routine. The bufr soundings did produce a dramatic change in the behavior of the mountain wave for two of the four events for which I had bufr soundings, but further investigation revealed spatial variations in the flow characteristics were more likely to be the source of this dramatic change then the improved vertical resolution (simulations were also generated using the grid point closest to the bufr sounding). On the other hand, the bufr soundings did produce stronger lee-slope velocities in all cases even when spatial variations were accounted for, which, in turn, suggests the vertical resolution of the initial profile may still be an important consideration. Conversely, these simulations suggest the point used to initialize the daily runs may not be the best choice. Returning to the historical windstorm cases, I found the overall performance of the two-dimensional model for these two upstream points was basically equivalent. In other words, the new point did produce forecasts that agreed with observations better than the forecasts produced by the original point on some occasions, but the reverse was also true.

The Eta grids available in the forecast office have a six-hour temporal resolution, whereas, some high wind events can be rather short-lived or sporadic. Hence, model forecasts generated using these data may come up short at times due to poor temporal resolution. On the other hand, the bufr soundings are available at one-hour temporal resolution. To test whether a higher temporal resolution might improve the performance of the model forecasts, simulations were generated using forecast soundings two hours on either side of 12 and 18 hr forecasts (f10, f14, f16, and f20). The results from these simulations suggest a six-hour temporal resolution is adequate to capture the response of the mountain wave to temporal changes in the large-scale flow.

As the number of runs for the Boulder setup increased, it became evident the false warning rate was unacceptably high (142 simulations between October 1996 and April 1997 produced wind speeds in excess of the high wind criterion when the corresponding observed wind speeds did not meet the criterion). The downstream extent of shooting high lee-slope flow varies, so I first considered whether moving the location from which the wind speed forecast is obtained farther downstream would significantly reduce the false warning rate. Moving the location 10 km downstream reduced the false warning rate by almost 40% (from 142 to 91). Although this revision did produce a significant reduction, 91 simulations in the false warning category is still unacceptably high. The version of the two-dimensional model used up to this point applied a free-slip bottom boundary condition, so the next logical step was to see if the addition of a simple surface friction parameterization would reduce the false warning rate to a more acceptable level. After adding a simple surface friction parameterization to the model code, I reran the simulations for the Boulder high wind events that occurred between December 1993 and April 1996 to determine the appropriate drag coefficient (i.e. a value small enough that actual high wind events are still forecasted by the model). Since requesting the Eta grids for all the false warning cases would have required too much disk space, I decided to use the corresponding bufr soundings. The archive I had generated contained soundings for 72 of the 91 cases under consideration, and 66 of these profiles also generated lee-slope velocities that exceeded the high wind criterion (false warnings) when the free-slip bottom boundary condition was applied in the model. With surface friction, the false warning rate was reduced by almost 60%. Although adding surface friction did significantly reduce the false warning rate, 28 false warning simulations in one season is still unacceptably high. At this point, I conferred with Brad, and he recommended I move onto testing the model in other regions, with the hope results for these regions might shed some light on the difficulties I had encountered while working on the Boulder region.

While working on the Boulder windstorms, I had also been busy compiling information on downslope windstorms that affected a number of other regions between 1993 and 1997, and preparing additional archive retrieval requests for these cases. Larry Dunn and Bill Alder (WSFO SLC) provided information on high wind events that occurred along the western slopes of the Wasatch Range, Carl Dierking (WSFO JNU) provided very detailed information on high wind events that occurred along the western slopes of Salisbury Ridge, David Bernhardt (WSFO GTF) provided information on high wind events that occurred along the eastern slopes of the Lewis Range, Kraig Gilkey and Eric Stevens (WSFO FAI) provided information on high wind events that occurred along the northern slopes of the Alaska Range, and Bob Clay (WSFO ANC) provided information on high wind events that occurred along the northwestern slopes of the Chugach Range. With the help of Brad Colman (WSFO SEA) and Brian Colle (University of Washington), I compiled a list of test cases for high wind events that affected Enumclaw, Washington (western slopes of the Cascades). I would also like to acknowledge the information Stephen Keighton (WSFO FLG) and Glenn Carrin (WSFO ABQ) sent me on high wind events that occurred in their CWAs. Due to time limitations and the limited number of high wind events I had to work with in these two regions (only one or two cases), I did not consider these regions viable test sites.

When I turned to testing the model in these regions, I returned to the free-slip bottom boundary condition, and a representative terrain profile was derived for each location following the same technique used to obtain a profile for the Boulder simulations. The projection angle used to determine the cross-barrier component of the large-scale flow is simply that corresponding to the direction perpendicular to the major mountain ridge axis. I also took a serious look at how I was initializing the model, and decided to take a different approach (there is some ambiguity on how to initialize the lowest levels and how the heights in the Eta profile correspond to heights in the two-dimensional model). The new approach for determining the height correspondence, which is far more logical than the approach I applied in the Boulder setup (not really worth describing at this point), simply subtracts the height reduction I applied to the representative terrain profile from the heights associated with the Eta profile. As for the initialization of the lowest levels, I first tried setting the surface wind speed to half that of the wind speed at the lowest level in the Eta profile (I did not use the surface values from the Eta model because I had been warned by various people these values were useless) in an effort to account for slowing of the low-level flow by frictional effects, but this approach produced lee-slope wind speed forecasts that were significantly weaker than the observed wind speeds for most of the Enumclaw, Salt Lake City, and Juneau high wind events. In an attempt to rectify this situation, I tried simply setting the surface wind speed to the wind speed at the lowest level in the Eta profile (i.e. a surface layer with a constant wind speed), which improved the results immensely. With this particular surface layer scheme, the forecast ability of the model for these three regions looks very promising. In fact, the model was able to capture the temporal evolution of the high wind events in these regions amazingly well on a number of occasions. The results for Anchorage, Fairbanks, and Great Falls using this setup were also promising, but the results suggested the model in its present configuration may have problems predicting the strength of high wind events in these regions. In light of my earlier work with Boulder, difficulties predicting the strength of the high winds also raises some concern of the potential for high false warning rates. On the other hand, the model setup used in the simulations for these regions is different from that used for Boulder, so it is difficult to say what bearing, if any, the results from my earlier work would have on the results for these regions. With these potential short-comings in mind, Brad and I decided the results for Enumclaw, Salt Lake City, Great Falls, Juneau, Anchorage, and Fairbanks looked promising enough to offer this forecast tool to these offices on a trial basis. By setting up the model to run in these local forecast offices, we will be able to more thoroughly evaluate how the model performs in these regions on a daily basis. We hope to hold another workshop on downslope windstorms in the Fall of 1999, at which representatives from each of these offices would be encouraged to present an overview of how the model performed in their CWA.

Although I had moved onto evaluating the model's performance in these other regions, I continued to look for an explanation for the high false warning rate associated with the real-time Boulder simulations. Using idealized modeling studies, researchers have shown strong downslope winds cannot reach the surface when a cold pool is present downstream of the mountain barrier. With these results in mind, I retrieved the Denver soundings from FSL's radiosonde database for the times corresponding to the 66 bufr sounding simulations with the free-slip bottom boundary condition that produced a false warning. A surface inversion was present in the Denver sounding for 88% of these cases. Of the simulations with surface friction that produced false warning (28), a surface inversion was not present in the Denver sounding for only 4 of these cases. Since the model is initialized with a horizontally homogeneous basic state based on a vertical profile obtained upstream of the mountain barrier, the model in its present form is not able to account for the presence of cold pools downstream of the mountain barrier, which, in turn, may account for a large percentage of the false warnings during this time period. This potential shortcoming of the model setup might be remedied by revising the method used to initialize the model (I have a few ideas I have not been able to investigate at this time) or by simply combining the model output with a decision tree that directs the forecaster to consider the conditions at the surface downstream of the mountain barrier.

The highlight of my COMET fellowship was my whirlwind tour of five forecast offices. I visited all three forecast offices in the Alaska region the last week of January, and then visited the Salt Lake City and Great Falls forecast offices the first week of February. During each office visit, I presented an overview of the dynamics of downslope windstorms, contrasted this type of terrain-induced high wind event with that of gap flow, discussed the potential interactions between these two mesoscale phenomena, described the setup of the model that would be running on a work station in their local forecast office, and presented the results for the historical high wind events in their CWA. This presentation took approximately an hour, give or take 30 minutes, depending on the audience (the forecasters were encouraged to ask questions and they did!). I gave two presentations at every office (an attempt to reach as many of the forecasters in each office as possible) except the Salt Lake City office, where I gave only one presentation (their choice). The remainder of the time during my visit was spent getting the model running on a workstation in the local forecast office, and working with the forecaster (or forecasters) who had volunteered to serve as the focal point for this new forecast tool. In addition, I gave two presentations at the Seattle forecast office and setup the model to run on a workstation in their office. The positive response I received from each forecast office was overwhelming, which more than compensated for the hectic travel schedule. I would also like to acknowledge all the work Carl Dierking put into adapting the input/output configuration of the model to the software used by the Alaska region. Getting the model to run on their system in time for my visit would not have been possible without Carl's help.

While visiting the forecast offices, the question of whether some type of surface friction parameterization might be needed was raised once again, and the question of whether moisture processes might be important for some locations was also discussed. In response to this feedback, I compiled instructions on how to use the surface friction and moist physics parameterization options that are available with the model code and sent these instructions to the focal point at each office that will be running the model.

During the second year of my fellowship, I gave talks on my research project at three workshops (Pacific Northwest Weather Workshop 1997 and 1998, Seattle, Washington; Intermountain Weather Workshop, Salt Lake City, Utah), and the American Meteorological Society's Conference on Weather Analysis and Forecasting (Phoenix, 1998). The response to my work at these meetings was also very positive. In fact, Dennis Gettman (WSFO Medford) and Keith Meier (WSFO Billings) approached me about the possibility of obtaining the model code so forecasters in their offices might use it as a learning tool. I hope to fulfill these requests sometime in the near future.

During my undergraduate and graduate studies, my exposure to the field of operational meteorology was rather limited. Hence, the time I spent interacting with the forecasters at the Seattle forecast office was a very important part of my fellowship experience. To familiarize myself with the forecast process, I observed lead forecasters during their shifts on a number of occasions. Working at the forecast office also gave me the opportunity to attend daily map briefings, weekly staff meetings, and various brown bag lunches and seminars, all of which gave me a better grasp of the operational environment. And finally, by working at the forecast office, I was able to share my academic training with the forecasters through one-on-one discussions and brown bag seminars related to terrain-induced flows. I found this interaction with the operational forecasters a very rewarding experience.

As always seems to be the case, there never is enough time to accomplish everything I would like to. I had hoped to submit an article on my project to the AMS journal Weather and Forecasting before moving onto my next position, a task I still hope to finish sometime in the next month or so (Brad Colman will be co-author for this article). I had also hoped to resolve the problems I encountered while trying to apply the model to the Boulder area. I will still be working toward a better understanding of the interaction between the large-scale flow and terrain in my current position, so I will continue to pursue the questions that were raised by my work with forecasting downslope windstorms along the Colorado Front Range. And finally, I had hoped to observe forecasters working the aviation and marine desks, and be trained to work at least the public service desk, but time simply ran out. The Denver forecast office and the laboratory I am currently affiliated with (ERL/ETL) will be moving into the same building in the near future, so I may still be able to pursue these interests in the future.