Prediction for Lateral Response of Monopiles: Deep Learning Model on Small Datasets Using Transfer Learning

This paper presents a novel approach for predicting the lateral capacity of large-diameter monopiles in multi-layered soil using deep learning and transfer learning techniques. With the increasing interest in offshore wind energy, the cost-effective design of offshore wind turbine foundations is crucial. Traditional methods such as the p-y method have limitations in analyzing larger diameter monopiles. As an alternative to high-fidelity numerical models (e.g., finite element analysis and finite volume analysis), deep learning models have gained popularity for load response predictions and design purposes. Yet, one major challenge of deep learning models is that insufficient data can result in poor model performance. In this study, a deep learning model incorporating convolutional and fully connected layers was developed to capture the complex interactions between pile geometry parameters and soil conditions. To address the challenge of limited dataset size, transfer learning was employed, leveraging a pretrained model on a large dataset. The results demonstrate that the proposed model can provide accurate predictions, even with a small dataset. Transfer learning significantly reduces the required data size while preserving high prediction accuracy.