Charitini Stavropoulou: Scalable modelling of marine renewable energy systems: From physics-based simulation to deep-learning surrogates

Abstract

Wave energy converters are commonly deployed in arrays forming wave energy conversion farms, particularly in the case of point absorbers, which represent a major class of wave energy technologies. Within such farms, devices experience strong hydrodynamic interactions through scattered and radiated waves on the free surface, making accurate modeling under realistic sea states highly challenging. Capturing these interaction effects is essential for reliable prediction of device motion, power absorption, and overall farm performance. However, high-fidelity hydrodynamic simulations of large arrays quickly become computationally prohibitive. This thesis addresses this challenge by developing a hierarchy of modeling strategies of increasing complexity, ranging from physics-based numerical formulations to multi-fidelity and fully data-driven spatial-temporal surrogate models, all aiming to achieve efficient and accurate farm-level predictions. The first part of the thesis presents a time-domain numerical model for single and multiple interacting devices based on linear potential flow theory. The governing equations of motion are formulated as convolution-type integro-differential equations, where radiation and wave excitation forces are expressed via impulse response functions. This approach enables simulation under irregular wave conditions while maintaining tractable computational cost. The model is validated against two wave tank experiments, demonstrating accurate reproduction of dominant motion characteristics and power absorption trends. It is further extended to larger configurations with several tens of devices, illustrating scalability for preliminary farm-level analysis. Building on this framework, a multi-fidelity learning strategy is introduced, combining low-order simulations with high-fidelity experimental data. A recurrent neural network is trained to learn the mapping between numerical predictions and experimental responses, effectively correcting systematic model discrepancies. This results in improved predictive accuracy while preserving computational efficiency, with enhanced robustness across varying wave conditions and array layouts. Finally, the thesis advances to fully data-driven spatial-temporal surrogate modeling, where array dynamics are inferred directly from experimental data. Both graph-based and transformer-based architectures are developed to capture coupled spatial and temporal interactions. These models demonstrate strong extrapolation capabilities, accurately predicting responses for unseen configurations, including incrementally expanded layouts. Overall, this work establishes a unified multi-level framework bridging physics-based and data-driven approaches for scalable wave farm modeling.