Efficient ultimate load estimation for offshore wind turbines using interpolating surrogate models L M M van den Bos, B Sanderse, L Blonk et al. Wind turbine site-specific load estimation using artificial neural networks calibrated by means of high-fidelity load simulations Abstract. Previous studies have suggested the use of reduced-order models calibrated by means of high-fidelity load simulations as means for computationally inexpensive wind turbine load assessments; the so far best performing surrogate modelling approach in terms of balance between accuracy and computational cost has been the polynomial chaos expansion (PCE). Regarding the growing interest in advanced machine learning applications, the potential of using Artificial Neural-Network (ANN) based surrogate models for improved simplified load assessment is investigated in this study. Different ANN model architectures have been evaluated and compared to other types of surrogate models (PCE and quadratic response surface). The results show that a feedforward neural network with two hidden layers and 11 neurons per layer, trained with the Levenberg Marquardt backpropagation algorithm is able to estimate blade root flapwise damage-equivalent loads (DEL) more accurately and faster than a PCE trained on the same data set. Further research will focus on further model improvements by applying different training techniques, as well as expanding the work with more load components. IntroductionTypically wind turbines are designed for specific wind conditions which are specified in site classes by the IEC standards. When a turbine is placed at locations where a site-specific parameter exceeds these design conditions, site-specific load assessments including simulations over the whole design load base have to be carried out. As this procedure can become computationally expensive, several methods and procedures have been developed for simplifying load assessments based on statistical moments, multivariate regression models [1] and expansions using orthogonal polynomial basis [2]. Previous investigations comparing different surrogate models such as polynomial chaos expansion (PCE), universal kriging with polynomial chaos basis function and quadratic response surface, have shown that the PCE results in the best overall performance for the load estimation in terms of robustness, accuracy and computing time [3].Regarding the growing interest in advanced machine learning applications, the purpose of this study is to evaluate the potential of using models based on Artificial Neural Networks (ANNs) as a flexible and potentially better-performance alternative to the previously mentioned surrogate models that have been developed already. Therefore, different ANN models are trained for

Recent studies have shown the advantage of replacing aeroelastic simulations with regression models based on Artificial Neural Networks (ANNs), which can be used as surrogate models for fast and efficient wind turbine load assessments. Once trained on a high-fidelity load simulation database covering a broad range of conditions, the surrogate model can be applied to predict loads for any site with wind climate falling within the range covered by the database. The aim of this study is to quantify the uncertainty propagation through such an ANN and to analyse how much the selected input variables influence the variance of the fatigue blade load estimations by means of a global sensitivity analysis. Results confirm that the selected ANN architecture seems suitable for this task resulting in small output uncertainties. Furthermore, the sensitivity analysis shows that the turbulence is mainly responsible for the blade load estimation, followed by the wind shear and the wind speed. The contributions of the turbulence length scale, turbulence anisotropy factor and wind veer angle are comparatively low. Comparing three different methods for sensitivity analysis shows that the partial derivative algorithm, Sobol variance decomposition and Shapley effect result in similar sensitivity measures.

This paper introduces a novel, transfer-learning-based approach to include physics into data-driven normal behavior monitoring models which are used for detecting turbine anomalies. For this purpose, a normal behavior model is pretrained on a large simulation database and is recalibrated on the available SCADA data via transfer learning. For two methods, a feed-forward artificial neural network (ANN) and an autoencoder, it is investigated under which conditions it can be helpful to include simulations into SCADA-based monitoring systems. The results show that when only one month of SCADA data is available, both the prediction accuracy as well as the prediction robustness of an ANN are significantly improved by adding physics constraints from a pretrained model. As the autoencoder reconstructs the power from itself, it is already able to accurately model the normal behavior power. Therefore, including simulations into the model does not improve its prediction performance and robustness significantly. The validation of the physics-informed ANN on one month of raw SCADA data shows that it is able to successfully detect a recorded blade angle anomaly with an improved precision due to fewer false positives compared to its purely SCADA data-based counterpart.

Abstract. In order to ensure structural reliability, wind turbine design is typically based on the assumption of gradual degradation of material properties (fatigue loading). Nevertheless, the relation between the wake-induced load exposure of turbines and the reliability of their major components has not been sufficiently well defined and demonstrated. This study suggests a methodology that makes it possible to correlate loads with reliability of turbines in wind farms in a computationally efficient way by combining physical modeling with machine learning. It can be used for estimating the current health state of a turbine and enables a more precise prediction of the “load budget”, i.e., the effect of load-induced degradation and faults on the operating costs of wind farms. The suggested approach is demonstrated on an offshore wind farm for comparing performance, loads and lifetime estimations against recorded main bearing failures from maintenance reports. The validation of the estimated power against the 10 min supervisory control and data acquisition (SCADA) power signals shows that the surrogate model is able to capture the power performance relatively well with a 1.5 % average error in the prediction of the annual energy production (AEP). It is found that turbines positioned at the border of the wind farm with a higher expected AEP are estimated to experience earlier main bearing failures. However, a clear connection between the load estimations and failure observations could not be confirmed in this study. Finally, the analysis stresses that more failure data are required in future work to enable statistically significant associations of the observed main bearing lifetimes with load exposures across the wind farm and to validate and generalize the suggested approach and its associated findings.

No abstract