Improving WAVEWATCH III hindcasts with machine learning: Fid-Bau Portal

Improving WAVEWATCH III hindcasts with machine learning

Lucero, Felipe / Stringari, Caio Eadi / Filipot, Jean-François

Abstract In this paper, machine learning models are used to improve a wave hindcast database created using WAVEWATCH III® (WW3) for the Chilean coast. The models were trained with 50,505 data entries from two buoys and eleven ADCPs. The machine learning models significantly improved the results from WW3 for three parameters: significant wave height, mean wave period, and mean wave direction. Our best performing model, which is based on a convolutional neural network and uses the directional wave spectrum as input, reduced root mean squared errors in the significant wave height by 71%, peak wave period by 61% and mean wave direction by 63%. Most importantly, our method dramatically improved the mean wave direction in four locations where WW3 was particularly problematic (absolute error reduction of 20°). The neural network corrections can also be applied to other locations if sea states conditions are similar to the training data. The research presented here show that machine learning techniques are a fast and effective way to improve existing wave hindcast databases at relatively low cost.

Highlights • Three machine learning models were tested to improve a wave hindcast database created using WAVEWATCH III® for the Chilean coast. • The models were trained with 50,505 data entries from two buoys and eleven ADCPs along the central south of Chile. • Our best performing model reduced RMSE: 71% in significant wave height, 61% in peak wave period and 63% in mean wave direction when considering test data only. • The neural network corrections can also be applied to other locations without increasing the errors in nodes that already had good agreement between WAVEWATCH III® and the observed data.