Abstract. The sustained development over the past decades of the offshore wind industry has seen older wind farms beginning to reach their design lifetime. This has led to a greater interest in wind turbine fatigue, the remaining useful lifetime and lifetime extensions. In an attempt to quantify the progression of fatigue life for offshore wind turbines, also referred to as a fatigue assessment, structural health monitoring (SHM) appears as a valuable contribution. Accurate information from a SHM system can enable informed decisions regarding lifetime extensions. Unfortunately direct measurement of fatigue loads typically revolves around the use of strain gauges, and the installation of strain gauges on all turbines of a given farm is generally not considered economically feasible. However, when we consider that great numbers of data, such as supervisory control and data acquisition (SCADA) and accelerometer data (of cheaper installation than strain gauges), are already being captured, these data might be used to circumvent the lack of direct measurements. It is then highly relevant to know what is the minimal sensor instrumentation required for a proper fatigue assessment. In order to determine this minimal instrumentation, a data-driven methodology is developed for real-world jacket-foundation offshore wind turbines (OWTs). In the current study the availability of high-frequency SCADA (1 Hz) and acceleration data (>1 Hz) as well as regular 10 min SCADA is taken as the starting point. Along these measured values, the current work also investigates the inclusion of an estimate of the quasi-static thrust load using the 1 s SCADA using an artificial neural network (ANN). After data collection all data are transformed to features on a 10 min interval (feature generation). When considering all possible variations a total of 430 features was obtained. To reduce the dimensionality of the problem this work performs a comparative analysis of feature selection algorithms. The features selected by each method are compared and related to the sensors to decide on the most cost-effective instrumentation of the OWT. The variables chosen by the best-performing feature selection algorithm then serve as the input for a second ANN, which estimates the tower fore–aft (FA) bending moment damage equivalent loads (DELs), a valuable metric closely related to fatigue. This approach can then be understood as a two-tier model: the first tier concerns itself with engineering and processing 10 min features, which will serve as an input for the second tier that estimates the DELs. It is this two-tier methodology that is used to assess the performance of eight realistic instrumentation setups (ranging from 10 min SCADA to 1 s SCADA, thrust load and dedicated tower SHM accelerometers). Amongst other findings, it was seen that accelerations are essential for the model's generalization. The best-performing instrumentation setup is looked at in greater depth, with validation results of the tower FA DEL ANN model showing an accuracy of around 1 % (MAE) for the training turbine and below 3 % for other turbines, with a slight underprediction of fatigue rates. Finally, the ANN DEL estimation model – based on two intermediate instrumentation setups (combinations of 1 s SCADA, thrust load, low quality accelerations) – is employed in a farm-wide setting, and the probable causes for outlier behaviour are investigated.

No abstract

No abstract

In this contribution SCADA data and thrust attained through strain measurements are used to train a neural network model which predicts the thrust load of an offshore wind turbine. The model is subsequently cross-validated for different turbines with SCADA data outside of the training period as input and the thrust load from strain measurements as the expected output, and the impact of wind speed and different operating conditions studied. The results for the model, such as MAE, are kept generally under 2 %. The estimated thrust load signal is then converted into the damage equivalent stress caused by the quasi-static load, allowing to quantify the damage induced by the thrust load. The model performed, in general, well, but some over-/underpredictions are severely amplified when converting the loads into the damage equivalent stress.

In this contribution we present a physics-informed neural network (PINN) approach for wind turbine fatigue estimation. This PINN incorporates physical information of the structure's fatigue profile in its loss function, referred to as Minkowski logarithmic error (MLE) -an extension of the log loss for any given L p space. The function is mathematically analysed and differentiated in order to better understand its behaviour. The results obtained using the MLE are favourably compared with previous efforts using the mean squared logarithmic error. Finally, the long-term error is evaluated based on the effect of p.