Towards Efficient Learning of Neural Fluid Force Estimators from CFD-Sampled Regular Waves

Abstract

Nowadays, ocean wave energy still represents a large, untapped, renewable energy source. To bring this potential to reality, the levelized cost of energy harvested by wave energy converters (WECs) must be reduced, to reach a viable commercialization of wave energy. This can be achieved by two means: by increasing the survivability and durability of WECs, and by improving their efficiency. The energy capture of WECs can be improved by using control, where controlled WECs can absorb significantly more power than uncontrolled, passive WECs. Most of the available control approaches are model-based, meaning that the performance of such control relies partly on their modeling accuracy. Paradoxically, control-oriented models must be computationally fast which often comes at the cost of prediction accuracy. The so-called linear potential flow approach is based on strong assumptions that limit the models to small wave amplitudes and displacement around the equilibrium position, greatly affecting the performance of the controllers based on such models.

In this work, we aim to bring new insights into the WEC-fluid interaction modeling, by modeling the total hydrodynamic force acting on the WEC using data-driven approaches. The training database will be provided by high-fidelity numerical simulations from computational fluid dynamics and the obtained non-linear surrogate model will predict the total hydrodynamic force acting on the WEC given the past WEC’ states and measurable fluid information. As such, the obtained model can be used within an online control framework. In addition, This project foresees the application of the implemented model to further improve the dry test rig developed at Uppsala University. By utilizing highly accurate hydrodynamics predicted by the surrogate model in real time, the buoy motion is precisely emulated while interacting with real physical power take-off units and controllers under testing. Particular attention will be given to the generalization capability and range of operation of the model which is a major concern in data-driven approaches.