Reliability-Based Design of Monopiles Using CPT Data and Deep Learning Enhanced Adaptive Metamodeling

Monopileis the mostcommontype of foundationsfor offshore wind turbines (OWT). A reliable design of monopiles requires considering the uncertainty in soil conditions. This study presents an enhanced framework for reliability-based design (RBD) of monopiles for OWTscombining a deep learning algorithm with a commonly used reliability technique called adaptive (or active learning) metamodeling. First, the soil vertical spatial variability was modelled using random field approach with the CPT data collected from the vicinity of the target installation location of monopile. A mechanical model of monopile was consequently created to supply target outputs for adeep learning algorithm used in this study called long short-term memory (LSTM) neural networks. LSTM predicts the outputs of the mechanical model to be used in the adaptive metamodeling for enhanced failure probability estimation against excessive monopile rotation.Following the failure probability estimations, RBD is performed using the proposed procedure to determine the minimum embedded length which satisfies a target reliability. The results disclosed that the deep learning enhancement to the active learning reliability estimated the failure probabilities effectively. The procedure proposed in this paper proves to be an enhanced tool for RBD of OWT monopiles. The objective of the future works is to develop an integrated algorithm for multiple RF inputs and optimize multiple design variables confirming a safe and cost-effective design.

Keywords: Monopile, Reliability-Based Design, LSTM, Active learning metamodeling, CPT data.