An innovative deep learning-based approach for significant wave height forecasting

Abstract

Significant wave height (SWH) is of critical importance in marine and coastal engineering applications and design, offshore operations, ship navigation safety, and other aspects. However, the complexity in generation and dissipation parts of wind wave fields makes accurate wave forecasting difficult. This study aims to develop an innovative deep learning approach based on combination of Variational Mode Decomposition (VMD) + Long Short-Term Memory (LSTM) + Transfer Learning (TL) for SWH forecasting. Firstly, transfer learning was applied to a ten-year data subset and different combinations of hidden and output layers were tried to be transferred for the analysis of the effectiveness of the transfer layer. Subsequently, the data is decomposed into Intrinsic Mode Functions (IMF) using the Variable Mode Decomposition (VMD) and given to the LSTM architecture. A model was proposed in which layer parameters are transferred serially between LSTM architectures for IMF's training. Eight different metrics such as mean squared error, mean absolute error, and root mean squared error etc. Were used to evaluate the performance of forecast models. Additionally, the model was tested using buoy wave measurements to demonstrate the effectiveness of the proposed method. The results show that the proposed method can successfully deal with nonlinear and irregular data structures. In particular, tests on buoy measurements have proven the method effective on real-world data. Besides, measurement-based forecasting using VMD + LSTM + TL model has highest correlation and lowest errors against the SWH measurements in comparison with SWAN-based forecasting using VMD + LSTM + TL model and wave forecasts estimated using the calibrated SWAN model forced based on the ECMWF IFS High-Resolution Operational Forecasts (ECMWF-HRES) wind forecast provided by the National Center for Atmospheric Research (NCAR).