Session: Government Agency Student Posters

Paper Number: 173097

Physics-Informed Gen-Ai With Sindy for Real-Time Prediction of Tower-Top Motions During Offshore Wind Turbine Installation 

The global expansion of offshore wind energy has led to the deployment of increasingly large turbines that deliver greater power output but introduce new challenges during installation. Among the most critical phases is the blade–hub mating process, where turbine blades are installed using crane vessels after the monopile, tower, and nacelle have been assembled in what is known as the hammerhead configuration. At this stage, the nacelle's center of gravity is eccentrically positioned relative to the tower axis, and wave-induced lateral motion causes torsional excitation and coupled dynamic behavior, resulting in complex orbital motions at the tower top. These motions pose a significant risk during blade installation, as excessive movement may lead to collisions at the blade root interface, safety hazards, and installation delays. Accurate real-time prediction of such motions is therefore crucial to support safe and efficient decision-making during offshore installations.

A generative AI-based digital twin architecture is currently being developed for this purpose, combining time-series encoding, autoencoder-based compression, and generative modeling in a latent space to simulate tower-top orbital motion under a range of installation scenarios. Orbital motion data (fore–aft and side–side nacelle displacements) are first transformed into 2D image representations using a reversible variant of the Gramian Angular Field (GAF), enabling consistent recovery of time-series signals from generated images. These images are then compressed into a latent space using an autoencoder, where a latent diffusion-based generative model operates to produce synthetic motion sequences conditioned on prior inputs.

However, the error in the predicted time-series motion and corresponding image representations is significantly larger for unstable orbital cases compared to stable ones, particularly near resonance conditions where the orbits change direction rapidly. These rapidly evolving dynamics pose a challenge for purely data-driven models, which often struggle to capture the underlying physical trends in such regimes. To enhance the reliability and physical interpretability of the generative model, a physics-informed extension is introduced by integrating Sparse Identification of Nonlinear Dynamics (SINDy) into the generative AI architecture. The SINDy framework identifies a low-dimensional, sparse ordinary differential equation (ODE) that captures the dominant physics governing tower-top motion. This identified ODE is incorporated into the model as an additional loss constraint during training, ensuring that the generated motion trajectories remain physically consistent while also improving convergence speed. By embedding physics-based structure directly into the learning process, the model benefits from both data-driven flexibility and physics-guided regularization, ultimately improving prediction accuracy, reducing training data requirements, and enhancing generalizability under varied sea state conditions.

Training data is generated using a high-fidelity numerical model of the IEA 10 MW reference turbine in the hammerhead configuration in OrcaFlex. The monopile and tower are modeled using Timoshenko beam elements that account for axial, shear, bending, and torsional dynamics, while the nacelle is represented as a rigid eccentric mass at the tower top. No blades or active controllers are included in the simulation. A range of irregular wave load cases is applied using multiple random wave seeds to generate diverse motion responses. Several thousand motion records are obtained, with 90% used for training and 10% reserved for validation.

Currently, high-fidelity datasets are being generated and tested to support model development and validation. Preliminary results show that the reversible GAF encoding retains full fidelity during conversion, and the autoencoder accurately reconstructs stable motion scenarios, performing robustly in low-noise conditions. Work is underway to integrate a physics-informed constraint using SINDy. The current focus is on identifying a sparse, low-dimensional representation of the governing dynamics, which will be used to enhance the generative model through additional physics-based loss terms. This ongoing development aims to improve model generalizability and interpretability by embedding structural knowledge of the coupled surge–sway–torsion behavior at the tower top. The long-term goal is to create a physics-informed generative digital twin that can support real-time prediction, planning, and risk assessment during offshore wind turbine installation.

Presenting Author: Saravanan Bhaskaran University of Maine

Presenting Author Biography: Saravanan is currently working as a doctoral student at Dr. Amrit Verma’s WEMO Lab at the University of Maine. The focus area of his PhD research is dynamic modelling and operability assessment of marine operations for offshore wind turbines. Saravanan research interests include simulation of marine operations, dynamic analysis using probabilistic methods and operation and maintenance of wind turbines. Before joining UMaine, he completed his master’s degree from the Norwegian University of Science and Technology (NTNU), Norway. During his master’s thesis, he worked on dynamic modelling and analysis of offshore wind turbine blade using jack-up crane vessels.

Authors:

Saravanan Bhaskaran University of Maine
Dyllon Dunton University of Maine
Yifeng Zhu University of Maine
Andrew Goupee University of Maine
Amrit Verma University of Maine