A regional wind wave prediction surrogate model based on CNN deep learning network. (September 2022)

• Record Type: Journal Article
• Authors: Huang, Limin; Jing, Yu; Chen, Hangyu; Zhang, Lu; Liu, Yuliang
• Is Part Of: Applied ocean research. Volume 126(2022)
• Publication Date: 2022-09
• Subjects: Regional wind wave prediction -- Surrogate model -- Deep learning network -- Convolutional neural network; Ocean engineering -- Periodicals; 620.416205
• DOI: 10.1016/j.apor.2022.103287
• Languages: English
• ISSNs: 0141-1187

Abstract

Abstract: Accurate prediction of wave characteristics of large-scale waves is of great significance for ocean disaster prevention and shipping safety. The traditional method is mainly based on ocean numerical models using wind as input to predict regional waves. In recent years, machine learning methods have gradually been concerned and applied in wave prediction, providing an effective approach for the real-time forecasting of regional wind waves. This paper proposes a regional wind wave prediction surrogate model based on a convolutional neural network (CNN), which takes historical wind and wave data as input to realize the prediction of current waves. Herein, the wind datasets from the European Center for Medium-Term Weather Forecast (ECMWF) and validated wave datasets simulated by the Simulating Waves Nearshore (SWAN) model are employed in model verification and optimization. We investigate the influence of historical winds and waves on the prediction accuracy of current waves and verify the adaptability of the surrogate model to different computational regions. The results show that the prediction accuracy is significantly promoted by considering historical wind and waves data in model input. The proposed model has good adaptability to different computational regions, and the results indicate that the surrogate model can effectively realize regional wind wave prediction, which greatly improves the computational efficiency as compared to the SWAN model. In addition, its relative error with the verified SWAN simulation results is less than 5%.