A data-driven approach leveraging multi-fidelity datasets for modeling arrays of wave energy converters

Abstract

To produce power of more than a few megawatts, wave energy converters are usually arranged in various geometric layouts or arrays, forming wave energy conversion farms. In particular, this is true for point-absorbers, the largest category of wave energy concepts. For any given farm configuration, the involved devices can experience intense hydrodynamic
interactions due to the scattered and radiated waves on the free surface; hence, predicting the overall system’s performance is a challenging task. The devices, as well as the extensive fluid domain associated with wave farms, limit the applicability of high-fidelity numerical methods due to high computational costs. Thus, low-fidelity methods, like linear potential flow theory, are typically used to model farm dynamics. However, the unrealistic
assumptions of the fluid flow can lead to rough approximations, especially in steep wave environments. To provide a more realistic scenario of the interacting behavior between the devices in a given layout, experimental wave tank tests could be employed. Because of the complexity and financial costs associated with such campaigns, experimental data for
wave energy conversion farms are scarce and often limited to a few layout configurations. Motivated by the intrinsic necessity of reliable models for performance estimation in wave farms, we present a multi-fidelity, data-driven method that employs low-fidelity numerical simulations and high-fidelity experimental estimates. Our model correlates a linear potential flow heaving estimate of several interacting point-absorbers arranged in various array layouts to the corresponding, high-fidelity, experimental, multi-degree-of-freedom motion of the devices. The method incorporates a long short-term memory neural network that predicts the devices’ temporal response to an incoming irregular wave within a farm layout that is unseen during the training phase of the model.