84th AMS Annual Meeting

Tuesday, 13 January 2004: 11:15 AM
Multimodel fine-resolution ensembles for short-range forecasts in mountainous terrain
Room 6A
Roland Stull, University of British Columbia, Vancouver, BC, Canada; and H. Modzelewski, X. Deng, Y. Zhou, L. Huang, T. Cannon, G. Hicks II, D. Storey, M. Holmes, and J. Charbonneau
Poster PDF (315.7 kB)
An operational, 4-model, NWP ensemble has been run daily for a year as part of a project to improve short-term, mesoscale, forecast skill over the complex (steep mountain, coastal) terrain of western N. America. The four models used in this study include NMS (from U. Wisconsin), MM5 (from NCAR and Penn State), MC2 (from RPN Canada), and WRF (from NCAR and NCEP). All the models are run for the same set of nested domains of 108, 36, 12, and 4 km horizontal grid spacing. In addition, the MC2 and MM5 are run at 2 km grid spacing, with this finest grid covering the Georgia Basin (Seattle, Victoria, Vancouver, and vicinity). All initializations are from the 00 UTC Eta analysis, from NCEP. Real-time forecasts and verification statistics may be viewed at http://weather.eos.ubc.ca/wxfcst/

Because of the steep topography, all the models have large biases due to the averaging of terrain elevation across individual grid cells. To correct for this, we have been testing Kalman-predictor post processing to remove the “localization” effects of terrain in our forecasts. From our daily verification results against roughly 500 surface weather stations during the whole year, we find that: (1) inclusion of the Kalman-corrected coarse-resolution grids with all the other medium and finer grids helps improve the ensemble average; (2) a simple average of the Kalman-filtered forecasts for all 18 ensemble members (from 4 models, each at 5 or 4 horizontal resolutions) performs equally as well as an average of the raw model outputs when weighted inversely by their error variances.

The procedures described in the previous paragraph have been applied to the following surface weather elements: temperature, humidity, wind speed and direction, pressure, and precipitation. We find that the equally weighted, Kalman filtered forecast makes improved forecasts for all these variables except precipitation. The Kalman filter-corrected precipitation forecasts usually verify worse than the raw model output. Experiments with neural network correction (which is a nonlinear approach, as opposed to the linear approach of the Kalman filter), performed no better than the Kalman filter; namely, it too made the forecasts worse for the mountainous terrain of British Columbia, Canada.

Given the very steep topography, we find that ensemble forecasts need both many members AND fine resolution (for at least some of the members). We also find, as was found by Mass and others, that finer resolutions often give worse verification scores. This has been traced to the lack of adequate in-situ wind and thermodynamic data in the lower and mid-troposphere over the Pacific Ocean, known locally as the Pacific Data Void. Further improvements in short range, 1-3 day, forecasts for British Columbia and Washington are unlikely until the data void is filled. The THORpex program aims to do just that, with a balanced mixture of in-situ and new satellite observations. A separate paper (Spagnol et al, 2004) at this conference presents our development of a Rocketsonde Buoy System, as one of the candidate in-situ systems for THORpex.

The human-machine interface challenge is being addressed with several new types of graphic displays for output products. We have found that Vis5D, while extremely useful in research mode, is too complicated and time-consuming for operational meteorologists, who must meet short deadlines. We are experimenting with a display called Ensuite, which combines the ease of viewing a computer animation, with that of a multi-panel chart. More importantly, this display experiment allows easy display of ensemble members. We have also been exploring other graphics products that are more useful for fine resolution in steep terrain than are the traditional synoptic-scale maps.

Supplementary URL: