83rd Annual

Monday, 10 February 2003
Applications of a roadway frost prediction system in Iowa
Bradley R. Temeyer, Iowa State University, Ames, IA; and W. A. Gallus Jr., E. S. Takle, T. M. Greenfield, and K. A. Jungbluth
Poster PDF (70.7 kB)
Each year, the state of Iowa has around thirty frost events on bridges and twenty frost events on roadways that cause hazardous travel conditions to the motoring public (Takle, 1990) . The Iowa Department of Transportation (IaDOT) spends approximately one million dollars per year on frost suppression activities to reduce the effect on travel conditions. Because of the expense incurred with frost suppression activities, it is valuable to have an accurate forecast of when frost conditions are present. Several different road frost, weather, and road temperature forecasts were examined and verified against human observations of frost on a bridge in Ames, Iowa during the winter of 2001-02. Frost occurrence, start and end time, as calculated by the frost deposition model were driven by meteorological conditions derived from various methods including (a.) a nearby Roadway Weather Information System (RWIS) station (within 10 km), (b.) forecasted by a mesoscale numerical model, (c.) forecasted by a private agency and (d.) forecasted by an artificial neural network. The RWIS observations represent as close to a perfect forecast as possible since these are observations used for the frost deposition forecast, where as the other methods are using various forecasting techniques.

The most complex of these forecasting techniques involves utilizing an artificial neural network to forecast the various parameters as needed by the frost model to make a frost deposition forecast at twenty minute intervals. To keep the complexity of the models as low as possible, a series of models have been developed to predict the input parameters needed by the frost model. The frost forecast is to be issued at 18 UTC daily, at twenty minute intervals beginning at 00 UTC and ending at 15 UTC. The neural network was trained on model output and observations for four different observation sites from three different cold seasons, 1995 1998 inclusive. It is believed that this method will provide more accurate forecasts of the various time-varying parameters to compute frost depth.

Supplementary URL: