Quantile Regression and Clustering Models of Prediction Intervals for Weather Forecasts: A Comparative Study

Information about forecast uncertainty is vital for optimal decision making in many domains that use weather forecasts. However, it is not available in the immediate output of deterministic numerical weather prediction systems. In this paper, we investigate several learning methods to train and evaluate prediction interval models of weather forecasts. The uncertainty models of weather predictions are trained from a database of historical forecasts/observations. They are developed to investigate prediction intervals of weather forecasts using various quantile regression methods as well as cluster-based probabilistic forecasts using fuzzy methods. To compare and verify probabilistic forecasts, a novel score is developed that accounts for sampling variation effects on forecast verification statistics. The impact of various feature sets and model parameters in forecast uncertainty modeling is also investigated. The results show superior performance of the non-linear quantile regression models in comparison with clustering methods.