Neural Networks to Derive Wave Spectra

摘要

Knowledge of a design wave spectrum is needed in works like structural analysis, and laboratory wave simulations. Use of theoretical equations due to Pierson-Muskowitz (PM) and JONSWAP is traditionally made for this purpose. This paper presents an alternative approach based on neural networks. Networks were trained using a variety of learning schemes in order to estimate shapes of the wave spectra from given values of the representative wave height and period. The validation of the network for unseen inputs showed that the neural network could be a viable option in order to estimate the shape of the wave spectrum from the specified design wave parameters. The network-predicted spectral shapes were more satisfactory than those yielded by the common theoretical spectra. While use of available wave time history could be much beneficial for training, the network can also reasonably learn from the theoretical spectra, albeit with some loss of accuracy.