Title:

The power of transfer learning in developing prediction model for offshore monopiles using limited data

The load response of large-diameter monopiles is highly nonlinear and site-dependent. The design of laterally-loaded piles has traditionally relied on the p-y method, which was originally developed for long, slender piles. Direct use of the p-y method for large-diameter monopiles (with small slenderness ratios) may lead to large errors in the estimated lateral pile capacity. Machine learning (ML) methods are superior in solving highly-nonlinear problems. In this paper, we developed a hybrid neural network model that predicts the lateral capacity of large-diameter monopiles based on the cone penetration test (CPT) data, pile geometries, and loading conditions. We constructed a hybrid neural network model by combining the convolutional neural network (CNN) and fully-connected (FC) deep learning neural network. The model is efficient in training and can generate high-accuracy predictions. A large number of synthesized data obtained based on rigorous finite element modeling were used to train the constructed hybrid neural network model. The developed DL model is able to predict the load-rotation response of monopiles in multi-layered sandy soil. The average relative error in the predicted lateral capacity is 3.1%.

Name

Comment*

Comments are viewable only by submitter

Captcha*