Application of machine learning to reduce operational wave forecast errors
A/Prof. Jeff Hansen1, Mr Kevin Vinsen2, Dr Stefan Zieger3, Dr Diana Greenslade3
1University of Western Australia, Crawley, Australia, 2International Centre for Radio Astronomy Research, Crawley, Australia, 3Bureau of Meteorology, Melbourne, Australia
Many marine industries require accurate wave forecasts to plan and execute operations both in coastal and offshore areas. While spectral wave models (e.g. WAVEWATCH III) have proven to be robust forecasting tools, and are numerically efficient, they are subject to errors and biases due to the input wind fields, parameterizations of physical process, and incorrect/under-resolved bathymetry. As a result, improvements in forecast skill can be difficult to achieve within existing operational modelling frameworks. Here we present results of a Machine Learning (ML) model developed to improve Bureau of Meteorology (BoM) wave forecasts at a port facility at Barrow Island in northwest Australia. Our approach was to train an ML model to identify and correct errors in a high-resolution regional wave model, which was nested within a global wave and atmospheric model. The ML model was trained using six years of archived spectral wave forecast data, spectral wave buoy observations, tide data and wind data. Following training, the ML model produced a revised 5-day forecast every 12 hours of the two-dimensional wave spectrum (energy as a function of frequency and direction) from which standard wave forecast parameters can be calculated (e.g. significant wave height). The ML forecast system was evaluated over the winter of 2022 (May-August) by computing the root mean square error (RMSE) of both the ML and BoM forecasts against the observed wave condition as recorded by a wave buoy. Over the evaluation period (193, 5-day forecasts) the ML model reduced the RMSE for the forecasted significant wave height by 6%, by 67% for the peak wave period, and 11% for the peak wave direction. For the peak wave period and direction, the error reductions were uniform across the 5-day forecasts whereas for the significant wave height the error reduction was the greatest in days 3-5 of the forecasts.
Biography:
Jeff Hansen is an Associate Professor at the University of Western Australia. His research expertise spans nearshore and coastal processes, surface wave dynamics, and marine renewable energy.
