Performance of Observation-Based Prediction Algorithms for Very Short-Range, Probabilistic Clear-Sky Condition Forecasting

Abstract

Very short-range sky condition forecasts are produced to support a variety of military, civil, and commercial activities. In this investigation, six advanced, observation (obs)-based prediction algorithms were developed and tested that generated probabilistic sky condition forecasts for 1-, 2-, 3-, 4-, and 5-h forecast intervals, for local and regional target types, in six geographic regions within the continental United States. Three of the methods were based on predictive learning algorithms including neural network, random forest, and regression tree. The other three methods were statistical techniques including a k?nearest neighbor algorithm, a classifier based on the Bayes decision rule, and a multialgorithm ensemble. The performances of these six algorithms were compared with forecasts from three benchmark methods: basic persistence, the climatological-expectancy-of-persistence, and satellite cloud climatology. The obs database for each forecast target was composed of a multiyear, half-hourly time series of atmospheric parameters that included cloud features extracted from weather satellite imagery and meteorological variables extracted or derived from data assimilation?based model analyses generated by NCEP?s Eta Data Assimilation System. The performances of the advanced prediction algorithms exceeded those of the benchmarks at all five forecast intervals for both target types in all regions, on the basis of a group of metrics that included receiver operating characteristic score, sharpness, accuracy, expected best cost, and reliability.