Seminar: Regime-switching forecast combinations: a 2-stage scheme for wind-farm energy outputs

AUEB STATISTICS SEMINAR SERIES MAY 2021

Yiannis Kamarianakis, Principal Researcher, Statistical Learning Lab, Institute of Applied and Computational Mathematics, FORTH

Regime-switching forecast combinations: a 2-stage scheme for wind-farm energy outputs

ABSTRACT

This work computes medium term forecasts (12-36 hours ahead) of wind farm energy production, using numerical weather prediction outputs. The first stage of the procedure evaluates alternative models, such as random forests, extreme gradient boosting, polynomial regressions with numerous predictors estimated with elastic-net and lad-lasso, in terms of their accuracy in, a) downscaling wind speeds at the wind farm locations, and b) forecasting energy production. In the second stage, selected energy production forecasts are combined, with weights that depend on the levels of forecasted wind speed. A regime-switching combination scheme based on a new Smooth Transition regression model is discussed.

(Presentation video can be found here)