Comparison of artificial intelligence and statistical techniques for probabilistic forecasting of sky condition

Cloud forecasting remains a significant challenge for meteorologists. Short-term, sky condition forecasts (0-6 hr) span a time frame where observation-based techniques typically give way to numerical modeling with regard to forecast skill. Statistical cloud forecasting techniques supporting aviation forecasting that utilize historical time-series, such as traditional stratified conditional climatology tables, based on surface observations date back more than fifty years. In this investigation, we compare the performance of seven, contemporary predictive learning (or classification) techniques in the generation of probabilistic forecasts of clear sky conditions for six different locations in the continental United States. The methods we applied from the fields of statistics, artificial intelligence and meteorology performed comparably well against a baseline of basic persistence. Our findings are relevant to a range of civil and military applications including aviation, reconnaissance, power load forecasting, and variational data assimilation.