Accelerating Offshore Windfarm Site Characterization Using Deep Learning

Abstract Developing offshore wind farms requires effective mapping of shallow subsurface for turbine foundation design, construction, and monitoring, all of which face many challenges in especially field data conditioning, structure interpretation and modeling, and geotechnical property estimation. In this paper, we revisit these challenges from the perspective of pattern recognition and propose implementing deep learning (DL) into automating three essential tasks in windfarm site characterization, including (i) cone-penetration testing (CPT) data conditioning, (ii) ultra-high resolution (UHR) seismic horizon picking, and (iii) integrated geotechnical property estimation, which leads to an accelerated workflow for delivering reliable ground models in an offshore windfarm site of interest. Specifically, the CPT data conditioning aims at identifying outliers in CPT data and reconstructing missing segments via a 1D auto-encoder. The UHR seismic horizon picking aims at tracking key horizons in collected UHR seismic via a two-step supervised DL and building a horizon model that captures the primary structural patterns in the target area. The integrated geotechnical property estimation aims at integrating the reconstructed CPT logs, the UHR seismic images, and the horizon models into simultaneously estimating multiple geotechnical properties such as cone-tip resistance (RES) and friction ratio (FRR) via semi-supervised DL. As tested over the public Borssele dataset within the Dutch Offshore Windfarm Zone, the proposed DL-accelerated workflow successfully improves the quality of CPT data, picks multiple major horizons that reflect the complexities of shallow subsurface, and constructs the corresponding RES and FRR models that delineate the lateral variations in geotechnical across the Borssele area.