Accurate wave prediction is essential for coastal and ocean engineering, as sea state conditions directly impact the design and operation of marine infrastructure, renewable energy systems, and maritime safety. While most research focuses on forecasting significant wave height (Hs) using increasingly complex models, other essential variables such as wave period (Tp) and direction (Dir) are often overlooked despite their importance in fully characterizing sea states.
This study addresses this gap by applying Artificial Intelligence (AI) models – Long Short-Term Memory (LSTM) networks and Random Forests (RF) – to predict Hs, Tp, and Dir. A novel window and flatten technique was introduced to restructure temporal data into a format suitable for machine learning, enhancing model performance for Dir and Tp predictions. Both models were tested under various wave conditions in the Mediterranean Sea
Results show that LSTM generally outperforms RF, particularly for Dir. However, RF models, which are not inherently designed for time series tasks, performed surprisingly well for Hs prediction and for short term Tp predictions. This opens promising avenues for developing hybrid models that combine sequential and non-sequential methods, potentially surpassing traditional sequence-to-sequence approaches in accuracy and robustness.
The study also highlights the challenge of accurately modelling Tp and the importance of evaluating model performance under varying energy conditions. Significant sensitivity to testing scenarios was observed, underlining the need for careful dataset selection and model validation. These findings provide a foundation for extending wave forecasting tools to more energetic environments such as the Atlantic Ocean and for advancing hybrid AI-based prediction frameworks.
扫码关注我们
求助内容:
应助结果提醒方式:
