In terms of applications, the focus of this research is on probabilistic short-term wind speed forecasting. The utility of non-deterministic predictions of future wind speed is well-known in the literature surrounding optimal wind turbine control and energy supply timing and management, as well as early-warning weather risk assessment systems. As successful meteorological tools exist for longer-term forecasts on the order of six hours or more, we consider forecast horizons of 1-6h and 1-10s, capturing key ranges required for a wide range of applications. We note that on the order of a few seconds, the literature on non-trivial probabilistic forecasters is exceedingly sparse, and even for the 1-6h range, successful models have been fine-tuned to detailed local conditions, and thus generally applicable methods have yet to be realized.
Our basic contribution is a method for deriving estimators defined in terms of loss functions with a property called propriety, and an empirical evaluation of the proposed approach against a collection of standard references. We note that while the desired propriety is in general model-dependent, we show for many important special cases that theoretically-motivated and computationally tractable estimators can be readily derived.
Experimental results show that the proposed forecasters uniformly outperform standard references across the prediction tasks considered, and are suggestive of a robustness to small deviations in predictive distribution that occur with a change of location. Our presentation highlights the key points of these results, and includes a discussion of the future lines of work that they suggest.