When can I trust my model? Data-centric machine learning model acceptability framework for farmwide lifetime assessment

Offshore wind turbine fatigue is more actual than ever. On the one hand, accurate estimated lifetime can serve as the basis for farm-life extension. On the other hand, for newer turbines with optimized design codes and reduced structural margins, overly aggressive control actions might jeopardize the fulfilment of their intended lifetime. To address this, recent years have seen a number of data-driven approaches, often based on artificial intelligence, that attempt to accurately estimate fatigue accumulation on offshore wind farms. The maturity of these approaches has steadily increased overtime, from initial single turbine-focused studies, to encompassing the entire farm and posing questions about its real-life, continuous employment. Thus, data-driven farm-wide fatigue assessment is now on the verge of industrial employment. This contribution further attempts to address real-world employment of data-driven models by posing a simple question: when can I trust my model? Specifically, it delineates a strategy - given the dynamic and unpredictable environment of continuous monitoring from varied data sources - to identify and deal with data unseen by the fatigue estimating model. This because, data-driven methods are generally great interpolators, but awful extrapolators, which means that predictions for out-of-distribution cases are highly questionable. For example, fatigue predictions during curtailment are only reliable if similar conditions have been seen during training or on the fleetleaders. If there is a a new curtailment set-point, the model might output very unrealistic values. We specifically focus on developing data-based (instead of model-based) tests to determine model confidence. The outcome is a framework which determines, given the incoming data, if the model fatigue predictions are to be trusted or put into question.