Mapping complex time series using LSTM neural networks: A case study on wave energy conversion farms

Often, point-absorber wave energy converters are arranged in farms. The fluid domain associated with wave farms limits the applicability of high-fidelity numerical methods due to infeasible computational costs. To model the motion within farms, linear potential flow theory is usually employed. However, the linear approximations lead to crude estimates, especially in steep waves. To capture accurately the motion of multiple devices, experimental tests are utilized. Due to the complexity associated, experimental data are rare and limited to few layouts and wave conditions. We propose a model based on long short-term memory neural networks that maps a linear potential flow heaving estimate of various interacting point-absorbers to the highfidelity experimental multi-degree-of-freedom motion of the devices. By learning the correlation between the low-fidelity input and the high-fidelity output, the model corrects the simplified heaving by a factor of two. The improved motion shows nonlinear behavior, initially only present in the experimental data.