Cluster-based regime-switching AR for the EEM 2017 Wind Power Forecasting Competition

This paper describes the regime-switching autoregressive models used to win the EEM 2017 Wind Power Forecasting Competition. The competition required participants to produce daily forecast wind power production for a portfolio of wind farms from 2 to 38 hours-ahead based on historic generation and numerical weather prediction analysis data only. The regimes used in the methodology presented are defined on the previous day's weather conditions using the k-medians clustering algorithm. Cross-validation is used to identify models with the best predictive power from a pool of candidate models. The final methodology produced a final weighted mean absolute error 4.5% lower than the second place team during the two-week competition period.