Abstract
An automated system for predicting maximum and minimum surface temperatures for 12- to 60-hr projections is described. The system uses multiple regression equations derived for 131 cities in the United States and 12 in southern Canada from 18 years of daily data stratified by 2-month periods. The predictors are selected by screening upper level heights and thicknesses observed at 67 grid points in North America and surface temperatures observed at the network of cities. On the average, about three-fourths of the temperature variance is explained by 4 or 5 variables, and the standard error of estimate is just over 4F. The system has been applied on an iterative basis twice daily at the National Meteorological Center (NMC) in Suitland, Md., since March 1963. Verification statistics are presented for 18 months of operational forecasts made by utilizing the barotropic and Reed numerical models as input to the multiple regression equations. During this period the automated temperature forecasts were superior to persistence and almost as good as subjective forecasts. Results of a one-month experiment are cited to demonstrate the improvement in temperature forecasting attainable by utilizing the NMC primitive equation model as numerical input to the system. Suggestions are also made for subjective improvements by considering factors neglected in the derivation.
