Data-driven short-crested wave prediction with improved generalization via hybrid feature standardization and training strategy

Accurate phase-resolved prediction of multidirectional short-crested waves is essential for offshore engineering applications, yet the performance of data-driven models often deteriorates when encountering unseen sea states because of the distribution shift between training and testing conditions. To improve cross-sea-state generalization, this study proposes a hybrid optimization strategy that combines feature standardization based on significant wave height and mean wave period with multi-sea-state data fusion. The proposed strategy is validated using wave-basin experiments covering four short-crested sea states generated under a Joint North Sea Wave Project spectrum. Four representative neural-network models, including long short-term memory (LSTM), temporal convolutional network (TCN), convolutional LSTM (ConvLSTM), and patch time series transformer (PatchTST), are employed to evaluate the model-independence of the proposed optimization strategy. The results demonstrate that the proposed strategy consistently improves the prediction accuracy of all evaluated models. Under the present experimental conditions, maximum root mean square error reductions of 37.2%, 19.0%, 43.7%, and 67.7% are achieved for the LSTM, TCN, ConvLSTM, and PatchTST models, respectively. These results demonstrate that the proposed optimization strategy effectively enhances the cross-sea-state generalization capability of the evaluated neural-network models under controlled laboratory conditions, providing a practical foundation for future validation under more diverse environmental conditions and field observations.