Non-homogeneous hidden Markov-switching models for wind time series

• Several Markov-switching autoregressive models are proposed for bivariate time series which describe wind speed and wind direction at a single location.
• The main originality of the proposed models is that the hidden Markov chain is not homogeneous, its evolution depending on past wind conditions.
• It is shown that the proposed models have good theoretical properties (ergodicity, consistency).
• It is shown that the proposed models permit to reproduce complex features of wind time series such as non-linear dynamics and multimodal marginal distributions.

In this paper we propose various Markov-switching autoregressive models for bivariate time series which describe wind conditions at a single location. The main originality of the proposed models is that the hidden Markov chain is not homogeneous, its evolution depending on past wind conditions. It is shown that they have good theoretical properties and permit to reproduce complex features of wind time series such as non-linear dynamics and multimodal marginal distributions. This is illustrated on a wind time series for a location off the French Atlantic coast using the R package NHMSAR.