Hidden Markov Models for Bounded, Inflated Time Series: Forecasting Icing on Wind Turbine Blades

Time series analysis of icing‐induced power loss in wind turbines pose several challenges: the response is bounded, serially dependent, intermittently missing, highly dispersed, and often inflated at a single value. We address these challenges with discrete‐time hidden Markov models for a discrete‐continuous process assumed to follow a mixture of state‐dependent zero‐inflated beta distributions. The framework allows covariates to influence either the transition probabilities or the distribution parameters. In a case study, we evaluate 12‐h‐ahead forecasts of icing‐related power loss. Compared with autoregressive and regression baselines, the proposed model achieves the highest predictive accuracy.