Wave Height Prediction Using Wind and Local Bathymetry with a CNN-LSTM model: A Case Study at Kyauk Phyu

Accurate wave prediction is crucial for coastal disaster management and maritime safety. Traditional numerical wave models, such as SWAN and WAVEWATCH III, provide physically reliable results but are computationally intensive and time-consuming. However, data-driven deep learning models offer fast prediction capabilities and have therefore received widespread attention in recent years. This study investigates regional wave prediction at Kyauk Phyu in the Bay of Bengal, an important coastal area undergoing ongoing port and infrastructure construction. Moreover, there are a limited number of wave buoys in the Bay of Bengal, which makes data collection difficult.      

In this study, the impact of hyperparameter optimization, input feature representation, and physically meaningful variables on the regional wave prediction is evaluated using a CNN-LSTM model. Hourly meteorological and wave data from ERA5 (2020–2023), and water depth information from GEBCO are employed in this study. The model’s hyperparameters are tuned using Bayesian optimization, and the result demonstrates that hyperparameter tuning plays a crucial role in spatiotemporal wave prediction. Subsequently, the performance of univariate and multivariate models is evaluated over different lead times of 1, 6, 12, 18, and 24 hours. The results show that the univariate model performs better for short-term predictions (1–6 hours), while the multivariate model incorporating wind stress and water depth achieves higher accuracy for long-term predictions (12–24 hours). This indicates that introducing more physical factors over a longer forecast period can enhance forecasting capabilities.

Performance evaluation during Cyclone Mocha shows that the model effectively captures high-energy wave events. An ablation method is applied to assess the contribution of additional features to wave-prediction performance. The results indicate that water depth is the most critical factor influencing wave-prediction accuracy, while the wind-stress variable results in only a slight change in prediction performance across all lead times.