New Improved Wave Hybrid Models for Short-Term Significant Wave Height Forecasting

In this paper, we have developed a series of wave hybrid models for significant wave height prediction. Our developed hybrid models uses a triplet of signal decomposition method, regression model and meta-heuristic algorithm. We have used the
$\epsilon $
-Support Vector Regression (
$\epsilon $
-SVR), Least Squares Support Vector Regression (LS-SVR), Long Short-Term Memory (LSTM) and Large-margin Distribution Machine based Regression (LDMR) model for the regression task. For signal decomposition methods, we have considered the Wavelet Decomposition (WD), Empirical Mode Decomposition (EMD) and Variational Mode Decomposition (VMD) method. Apart from this, we have also used the Particle Swarm Optimization (PSO) method to tune the parameters of the used regression model in our wave hybrid models. Till now, the VMD method and LDMR model have not been used in any wave hybrid model. We have evaluated the performance of our developed wave hybrid models on time-series significant wave heights, collected from four different buoys using the different evaluation criteria. After the detailed statistical analysis of the obtained numerical results, we conclude that the VMD-PSO-LDMR based wave hybrid model obtain best performance on six datasets out of seven considered datasets. Also, the VMD based wave hybrid models can obtain better performance than other decomposition based hybrid models. Further, we also conclude from our numerical results that the LSTM model outperforms the SVR, LS-SVR and LDMR based hybrid models if we do not decompose the significant wave height signals apriori. But, when we decompose the SWH time-series signals using a particular decomposition method, then SVR, LS-SVR and LDMR based hybrid models tend to improve their prediction ability significantly.