A machine learning based approach for better prediction of fatigue life of offshore wind turbine foundations using smaller datasizes
Abstract. As offshore wind turbine (OWT) foundations approach the end of their design life, the industry is increasingly focused on strategies for lifetime extension. As fatigue is the design driver for foundations of OWTs, reliable fatigue damage predictions are essential to support informed decisions for lifetime extensions. While simulation-based fatigue life reassessments are common, data-driven approaches using measured strain data have emerged as an alternative that can reduce modeling uncertainties. But, data-driven approaches face challenge as having access to strain data over the entire past lifetime is not an industry-standard. Often measurement campaigns are only kicked off when a lifetime extension is considered, thus limiting the availability of strain data. However, environmental and operational conditions (EOCs) of the wind turbines are usually recorded during the whole operational period. Using limited strain measurements and long term EOCs to estimate fatigue damage in unmonitored periods during the lifetime of the turbine requires temporal extrapolation techniques. Existing work on this topic presents several extrapolation methods, including linear time-based extrapolation, binning based on correlations between EOCs and average damage, and machine learning (ML) models. The accuracy of these methods depends on factors such as the selected EOC parameters, the duration and starting point of available strain data, the power rating and type of the wind turbine, as well as the type and architecture of the extrapolation model used. This study presents a novel machine learning based extrapolation model using random forest (RF) for temporal extrapolation of strain measurements. A comparative analysis of novel RF model with previously identified binning models is presented. The extrapolation performance is validated using five years of measured strain, SCADA, and wave data from a 3 Mega Watt (MW) and a 9MW OWT installed on monopile foundations in the Belgian North Sea. Using a sliding window approach on the available monitoring data, we estimate and compare the statistical uncertainty in fatigue life predictions of various extrapolation models. The results indicate that wave parameters play a more significant role in fatigue prediction for larger turbine of 9MW compared to smaller one of 3MW power rating. For limited data sizes, less than 12 months, the proposed RF model demonstrates superior performance, offering more reliable fatigue life predictions with reduced statistical uncertainty. However, for longer datasets, greater than 12 months, the performance advantage of RF model over binning methods becomes less pronounced.