Truncated Gaussian Hidden Markov Models and Wind Speed Spatio-Temporal Series

This paper discusses the use on wind speed data from NREL of a hidden Markov model specially crafted to reproduce the features of hourly multi-site wind speed time series, and to forecast them. Differently from the usual conditional multivariate Gaussian HMM, in this model the emissions are chosen as conditional multivariate truncated Gaussian distributions, which have non-negative support and are thus more suitable for non-negative data like wind speeds.