Neural-Network-Based Data Assimilation to Improve Numerical Ocean Wave Forecast

This paper demonstrates the skill level of a wavelet neural network in improving numerical ocean wave predictions of significant wave height (H
₈
) and peak wave period (T
_p
) having practical applications in operational centers. The study uses data of H
₈
and T
_p
for a coastal region off Puducherry located in the east coast of India, and obtained from a high-resolution wave model resulting from nesting of the SWAN model with the WW3 model. A wave rider buoy located off Puducherry provided data for a period of 25 months during the period from June 2007 until July 2009 used in this study. The time series of error between numerical and corresponding measured values was first constructed, and using a wavelet neural network, the errors were predicted for future time steps. The predicted errors when incorporated into the model values provided the updated prediction of H
₈
and T
_p
. The study signifies that numerical estimations could be significantly improved using this procedure. The results provide quite satisfactory predictions with a lead time varying from 3 to 24 h. The study points out that adequate training of the neural network is an essential prerequisite to obtain good performance and skill levels. A comparison between the suggested prediction method with the standalone neural network model trained with measured data off Puducherry showed that the former approach is preferred over the latter in obtaining a sustained prediction performance.