An ensemble strategy for road weather applications

Ensemble modeling methods for supporting road weather operations are described as they were applied in the first three years of a road weather field project called MDSS (Maintenance Decision Support System).

In general, an ensemble modeling system needs to include ensemble members that represent the natural variability of the targeted weather. Ideally, ensemble members should each be equally likely of being the best-scoring predictor for any given event, and their forecast errors uncorrelated. Ways of achieving dispersion among an ensemble set include varying the lateral boundaries (of limited-area models such as those used here), using various models or physical parameterizations therein, and using a variety of initializations.

MDSS is a project aimed at improving weather services for surface transportation, specifically winter weather operations. The MDSS system takes weather forecast information from many different sources, optimizes predictions for a few dozen points along highways where there are observations sites for validation, and produces suggestions to snow plow supervisors regarding best places to plow, how often, and how much salt or other chemical to use. This paper reports on MDSS demonstrations in central Iowa during two-month periods beginning in January 2003 and December 2003.

It will be shown that for this application, varying the lateral boundary conditions added negligible dispersion to the ensemble. On the third iteration, we settled on two different models (MM5 and WRF), 24-hr model runs on identical grids, reinitializing each model every hour with diabatic initialization grids produced by FSL's Local Analysis and Prediction System relying heavily on data from GOES satellites and WSR-88D radars. Time-lagged methods are used in the ensemble post-processing, e.g., a 5-hr forecast would use 5-hr, 6-hr, and 7-hr model outputs, all valid at the same time, from each of the two models, for a total of 6 members in the ensemble.

Verification of the models' precipitation forecasts, as well as emerging requirements from the client, guided the evolution of the MDSS model ensemble. Emphasis has evolved toward optimizing the very short-range predictions, in the 2-8 hr range. This has led to careful scrutiny of the timing of data arrival, the efficiency of the assimilation and initialization codes, and model run scheduling on a rather small (current replacement value about $35K) multiprocessor Linux-based compute platform. The payoff is 1-hr forecasts available to forecasters and post-processing software before those 1-hr forecasts are valid, e.g., the 1-hr forecasts from the 1200 UTC model runs are out before 1300 UTC.