Prediction of near-field uni-directional and multi-directional random waves from far-field measurements with artificial neural networks

Abstract

Predictions of several minutes of phase-resolved waves ahead of time from the measurements of wave-time-series at upstream locations are important for many marine operations. Making use of the knowledge of wave propagation, we design the input-output data and fine-tune the architecture of an artificial neural network (ANN) to push its capacity for the wave predictions. The study cases include uni-directional and multi-directional waves with mild to large wave steepness. The ANN models are trained on simulated random waves of JONSWAP spectrum with significant wave height and peak spectral wave period taken from a North Sea wave statistical data. Results demonstrate that the ANN models with single or multi-inputs from a 1 km upstream far-field locations can predict near-field waves with errors ranging from 2.6% to 12% for both uni- and multi-directional waves. The model accuracy is dependent on the wave steepness. Sensitivity studies show that the network model's hyper-parameters might need to be changed for different wave conditions. There are also possible optimal locations for far-field wave probes that will give the optimal prediction at the near-field.