ENSEMBLE WAVE FORECASTS BY DEEP LEARNING BASED ON APPROXIMATE BAYESIAN ESTIMATION

Abstract

 Short-term and medium-term wave forecasting is essential for process control in offshore construction projects up to one or two weeks in advance. In addition to physical models, statistical models, such as deep learning models, have recently been used for wave forecasting, and weather forecast values are used to accurately forecast waves several days in advance. However, the accuracy of wave forecasting decreases with an increase in forecast lead time due to the uncertainty of weather forecast values, making it challenging to accurately forecast waves one week ahead or longer. In this study, ensemble wave forecasting was conducted using Bayesian deep learning based on approximate Bayesian estimation, that takes into account uncertainties in weather forecast values and model representations. Compared to the deterministic wave forecast of the existing model, the 11-day wave forecasts of the proposed model improved the forecast accuracy by an average of 10 % ~ 20 %.