Probabilistic Virtual Load Monitoring of Offshore Wind Substructures: A Supervised Learning Approach

In this work, a virtual load monitoring framework is proposed for deriving a mapping from either high or low frequency (1-Hz /10-minute time averaged) SCADA data to load signals, while preserving the high frequency dynamic components of the latter. Specifically, the proposed virtual load monitoring scheme relies on a data-driven model that receives features retrieved from SCADA data and yields the probability distribution of the structural response. The constituent neural networks are trained via supervised learning based on the labelled data retrieved while strain sensors are still functional, since at that operational stage, both SCADA and structural response can be collected concurrently.
Once the strain sensors are not functional, the trained deep neural network is deployed, providing structural response predictions from on site SCADA data. The proposed virtual monitoring approach is tested on a monopile-supported offshore wind turbine and cross-validated in terms of the predicted stress range distribution of a structural connection
located at the mudline. The results show good agreement between structural response predictions and measurements, thus demonstrating the efficacy and utility of the tested scheme.