Using machine learning to improve operational wave forecasts

Operational wave forecasts rely on spectral wave models that due to their numerical implementation (i.e. phase-averaged) and resolution, either parametrize or do not fully resolve key physical processes that impact wave generation, propagation, and dispersion. These factors, coupled with potential errors in atmospheric forcing, can sometimes result in incorrect forecasts for wave conditions and/or the timing of their onset. Many offshore industries depend on accurate wave forecasting, and unexpected conditions may incur cost (due to halting an underway operation or a missed opportunity to complete an operation) or add safety concerns. In this presentation we outline results from an initial study to test the use of machine learning to adjust Australian Bureau of Meteorology AUSWAVE-R wave forecasts. Two years of archived wave forecasts, each extending 72 hours, were extracted at the location of three Western Australia Department of Transport directional wave buoys. Eighty percent of the observed and forecast wave conditions were used as a training data set for a Recurrent Neural Network algorithm which was then used to adjust the remaining 20% (randomly selected and independent from training data). This initial test resulted in the root mean square error of the forecasts being reduced by one-third for significant wave height and by nearly one-half for peak wave period and direction across all sites. Currently the technique is also being applied to the spectral data from the buoys and forecasts. These initial results indicate that machine learning can be an effective mean of improving existing operational wave forecasts with negligible additional computation.