Mitigating domain shift for cross-station wave height prediction in South China Sea island and reef waters: proposal and systematic evaluation of three novel methods

To address the domain shift problem in cross-station significant wave height prediction in South China Sea island and reef waters, this paper systematically proposes and evaluates three novel domain shift mitigation methods. Under the unified framework of a single-factor LSTM predictor, an evaluation framework encompassing "overall error distribution-single station accuracy-computational efficiency" was constructed to systematically compare the baseline method, the proposed "heavyweight" adaptation methods BO-FTTA (Bayesian Optimization Full-model Test-time Adaptation), BO-FBAT (Bayesian Optimization Freezing BN-layer Adaptive Test-time Tuning) and innovative "lightweight" adaptation method TDNA (Target Domain Normalization Adaptation). The study finds that all three methods significantly enhance cross-station prediction accuracy, but with markedly different computational overhead. Compared with the baseline method, BO-FTTA and BO-FBAT reduce the overall error standard deviation by 52%–53%, and achieve more than 50% NMAE reduction in high-discrepancy cross-station scenarios. In contrast, TDNA obtains comparable performance improvement, with its calculation time merely 0.9–1.2 times that of the baseline method, keeping similar computational efficiency. Accordingly, the research clarifies the applicable scenarios for different methods: the "heavyweight" optimization methods are suitable for offline high-precision optimization, while the "lightweight" adaptation method provides a feasible solution for online, real-time deployment scenarios. This study provides quantitative references for method selection in cross-station wave height prediction of South China Sea island and reef waters, and also offers valuable insights for other marine time-series cross-domain adaptation issues.