On machine learning and statistical data-driven models to improve wind farm reliability and operational condition forecasting

This PhD thesis contributes to advances in the modelling and management of wind farm reliability by integrating data-driven statistical and machine learning approaches for wind time series forecasting, along with a novel Bayesian approach for estimating wind turbine failure rates. Beginning with a systematic assessment of reliability fundamentals and empirical evidence, the work identifies the persistent challenges associated with critical wind turbine components. Building on this, state-of-the-art deep learning architectures are explored for both short-term wind speed and power forecasting, demonstrating that hybrid and explainable models provide both predictive accuracy and interpretability. Further developments extend to multi-horizon forecasting under data-sparse conditions, where recent neural methods are shown to outperform conventional baselines. Complementing these advances in prediction, a hierarchical Bayesian framework is introduced for estimating turbine failure rates under scarce and heterogeneous field data. By pooling information across sites, this approach yields stable estimates, quantifies uncertainty, and mitigates the limitations of maximum likelihood methods. Together, the contributions presented in this thesis provide a coherent framework that improves wind farm operational simulations by coupling accurate forecasting with principled reliability modelling, thereby supporting improved availability, cost-effectiveness and resilience in large-scale wind energy systems.