Session: 02-16-01 Data-driven Models for Marine Structures 1

Paper Number: 81048

81048 - Artificial Neural Network Surrogate Modeling for Offshore Wind Turbines Under Multi-Hazards 

Seismic safety concerns have increased on monopile-supported offshore wind turbines (OWTs), one of the prevailing marine structures for renewable energy supply. Since OWTs are persistently subjected to wind and wave loadings at work, multi-hazard effects need to be addressed in the seismic safety assessment for OWTs. It is challenging, however, in integrating such complex loading scenarios into the modeling framework for the soil-monopile-tower-rotor system where high levels of complexities and uncertainties are inherent. Recently, data-driven surrogate modeling has been found a viable solution for this problem. In particular, the multi-layer neural network (MLNN) is a promising machine learning method to derive relationships between the characteristics of the structure–load system and structural responses given datasets without physics-based transient simulations which require high computational costs. Yet, how the architecture of MLNN (e.g., numbers of hidden layers and neurons) and its impact on the model accuracy are not well understood. Therefore, this study is to develop a reliable NLNN-based surrogate model against earthquakes for monopile-supported OWTs in clay deposits under combined wind and wave loads.

A coupled soil-pile finite element modeling technique is first validated using a centrifuge test for the monopile-supported OWTs under wind, wave, and earthquake loadings. To prepare an adequate dataset, a series of dynamic analyses are then performed considering various uncertainties in soil, structural, and loading properties. In NLNN, these properties are used as input variables to yield the maximum bending moment of the OTW model as the output variable. The MLNN architecture is optimized in a probabilistic manner through a multi-fold cross-validation analysis. Improved understanding of the relationship among the number of hidden layers and neurons with respect to different input and output variables will play a pivotal role in the rapid design and lifetime management of monopile-supported OWTs under multiple hazards using reliable surrogate models.

Presenting Author: YeongAe Heo Case Western Reserve University

Authors:

Xiaowei Wang Case Western Reserve University
YeongAe Heo Case Western Reserve University