Propagating ml model uncertainties in offshore wind turbine fatigue damage accumulation

Modern offshore wind farm operators are increasingly looking at keeping tabs of
the structural lifetime consumption of their assets. As lifetime estimation is closely linked with complex decision-making processes like lifetime extension or certification, there has been a move from a deterministic framework to a modeling which quantifies uncertainties. This is especially relevant during unseen load- or new environmental conditions, where uncertainty quantification can bring light into understanding the severity of these phenomena and how to appropriately deal under uncertain conditions (e. G. Discard prediction, retrain model). This has led to data-driven approaches with probabilistic reasoning for fatigue load estimation. In this contribution, bayesian neural networks (bnn) and conformal predictors are compared for fatigue load prediction and uncertainty quantification. These outputs have a ten-minute
frequency (industry standard). However, fatigue accumulates non-linearly over time. Thus, ten-minute estimations need to be rescaled and aggregated, along with their corresponding uncertainties to reflect the entire interval under study and, eventually, the lifetime. As such, in this contribution, we attempt to extend this approach by devising a methodology to aggregate and propagate the model uncertainty related to each ten-minute prediction over a one-year
period of real-world data from an offshore wind turbine through Monte Carlo simulations and discuss its impact for lifetime.