Systematization of short-term forecasts of regional wave heights using a machine learning technique and long-term wave hindcast

Abstract

Machine-learning techniques have been applied to wave forecasting to address the implementational and computational challenges of conventional numerical wave models. Most previous studies emphasized the accuracy of approaches based on in situ wave measurements, but such approaches are site-specific and lack the spatial and temporal fidelity required by regional wave climate forecasts. This study presents a novel systematization of forecasting regional wave heights at sufficient spatial resolution and coverage, using long-term wave hindcasts generated by numerical wave models as a training dataset for a long short-term memory (LSTM). This systematization not only ensures the accuracy of the model's forecasts but demonstrates the applicability of long-term wave hindcasts as input data. It involves thoroughly investigating the dependence of model accuracy on multi-task LSTM structures (training variables, input data record periods, and forecast lead times), and on regional wave climates (different wave sites and sea states). The resulting model performance is spatially heterogeneous in global wave forecasts, underlining the need for validation of wave forecasting models to global wave climates. This study reveals that machine-learning techniques that leverage validated long-term wave hindcasts can successfully forecast regional wave climates. As access to high-resolution regional wave hindcast data improves, accuracy will also increase.