A Statistical Approach to the Short-Period Prediction of Surface Winds

Abstract

The screening-multiple-regression technique is applied to predicting surface u- and v-wind components at Idlewild International Airport for periods of 2, 3, 5 and 7 hr. The predictors are variables from 11 synoptic stations, easily obtained or derivable from conventional service A teletype data. Additional predictors are used to account for diurnal and seasonal variations. In all, 141 predictors are screened and one prediction equation is obtained for each predictand. Each equation is applicable to any hour of the day and any day of the year. The regression equations derived from a dependent sample selected randomly from 7 years of data proved significantly better at the 1-per cent level than both persistence and climatology for the 3-, 5- and 7-hr forecasts and at the 5 per cent level for the 2-hr forecasts when tested on 1387 independent cases. The screening-regression root-mean-square errors on this independent set ranged from 3.36 kt to 4.48 kt for the u-wind forecasts and from 3.69 kt to 5.57 kt for the v-wind forecasts. Operational 3-, 5- and 7-hr surface-wind forecasts extracted from terminal forecasts made at Idlewild are compared both quantitatively and categorically with corresponding regression forecasts made on a new set of independent data. The screening-regression forecast errors are approximately ? smaller than the subjective errors, and the improvements for all the predictands are statistically significant beyond the 1 per cent level. The categorical comparison concerning only categories of <10 kt and ≥10 kt (dictated by the format of the subjective data) resulted in Heidke skill scores of 0.399 for screening regression and 0.249 for the subjective forecasts when applied to 7-hr prediction of the surface-wind speed at Idlewild.