Probabilistic forecasts of wind speed using the bootstrapped Markov regime switching model

Abstract

With a recent surge of different alternative fuels, wind power has proven to be one of the most attractive option. In order to harvest its power, accurate and timely short-term forecasts of wind speeds are needed. We introduce the Bootstrapped Markov Regime Switching (BMRS) model to produce fully probabilistic 2-hour ahead forecasts of wind speed. Our model building strategy consists of 2 elements: The first is to explore the possibility of producing competitive Bayesian probabilistic forecasts using only on-site wind information as opposed to spatial-temporal approaches. Secondly, we hope to eliminate the need to specify beforehand a parametric predictive distribution and adopt bootstrap methodologies to produce probabilistic forecasts. By coding 90 degrees to 270 degrees clockwise as West and 90 degrees to 270 degrees anticlockwise as East, we assume wind speed is driven by two different autoregressive processes, each corresponding to a different regime of wind direction (East or West), and the transition from one regime to another is governed by a two state Markov chain. In this paper, we introduce the circular Markovian local bootstrap technique to model future state evolution. The parameters are estimated using a modified Gibbs sampler, while the final distribution-free predictive intervals for future wind speeds are developed using sieve bootstrap. The proposed method is applied to real life wind speed data in a case study.