A significant application of artificial intelligence to environmental science is the use of numerical weather prediction (NWP) models to forecast future weather. Operational NWP models regularly provide forecasters with output guiding them toward determining the evolution of the state of the atmosphere. However, when presented with an NWP model forecast, in addition to the predicted values of meteorological variables it would be desirable to have an idea of the potential for error in the forecast. This information, if accurate, would prove to be very useful to both researchers involved in the development of NWP models and operational forecasters responsible for issuing forecasts to the public.

Three-hourly predictions of the errors in 2-m surface temperature forecasts generated by the Advanced Regional Prediction System (ARPS) will be made for a large number (150-200) of surface observing stations in the continental United States. These predictions will be made by using a range of statistical techniques including Markov chains, non-linear analysis, regression, and error recycling. The error forecasts will be measured against NCEP model output statistics (MOS) temperature forecasts for the chosen sites to determine skill. Optimal linear combinations of these forecast schemes will also be examined. In order to form a statistical basis on which to use these methods, hourly surface data from each of the sites and archived ARPS forecasts will be collected. Error forecasts will be made for both dependent and independent NWP model data sets. Errors in temperature forecasts initially will be projected out to 4 days (96 hr) from model initialization. These projected errors will give a quantitative measure of confidence in the forecast at each site every 3 hours.

To view some of the results produced by these methods as they become available, please visit the following URL:

http://weather.ou.edu/~aataylor/research