CRBM-DBN-based prediction effects inter-comparison for significant wave height with different patterns

Based on the Conditional Restricted Boltzmann Machine - Deep Belief Network (CRBM-DBN), we select four patterns and compare their prediction effects for the significant wave height in the Gulf of Mexico (GoM). Historical datasets of all 12 buoys managed by the National Data Buoy Center are employed to train and construct models. Root-mean-square error (RMSE) and coefficient of efficiency (CE) between the observed and predicted wave heights are investigated. We find that for the short-term prediction (i.e., lead time≤12 h), the best results (RMSE 0.92) are achieved with the univariate significant wave height as the input in most cases of the whole gulf. When the lead time is equal to 24 h or 48 h, the multivariate pattern of “significant wave height + dominant wave direction + wind speed + wind direction” has the optimal effects (0.18 m