Multivariate GRU and LSTM models for wave forecasting and hindcasting in the southern Caspian Sea

Reliable wave forecasts and hindcasts are pivotal elements for a wide range of marine activities, coastal engineering applications, renewable energy and wave climate studies. This study explores capability of recurrent neural networks for wave hindcasting and multi-step wave forecasting in the southern Caspian Sea, a region with a limited observational data. Gated recurrent unit (GRU) and long short-term memory (LSTM) models are fed by buoy records for wave forecasting up to 6 h ahead and ERA5 reanalysis data as the model inputs for wave hindcasting. Special efforts are devoted to improve the accuracy of the models by including some additional features (e.g., wave age and swell) in the input layer. Different error measures are employed to evaluate performance of the models at two stations, Amirabad and Anzali. Using multivariate GRU and LSTM models, suitable wave forecasts and hindcasts are achieved for the locations. Although the model performance deteriorates with extending the forecasting time horizon, the forecasts of significant wave height (Hs) up to 6 h ahead are still satisfactory. The best hindcasting model yields a high linear correlation between hindcasted and measured Hs with R2 = 0.75 for Anzali and R2 = 0.83 for Amirabad. Finally, wave age and significant height of total swell (Hs-swell) can be considered as tentative input variables to improve efficiency of the models.