E-Thesis 318 views
Forecasting Nearshore Wave Conditions With Machine Learning Methods / MAHSA MOKHTARI
Swansea University Author: MAHSA MOKHTARI
DOI (Published version): 10.23889/SUThesis.70078
Abstract
Accurate wave height prediction is crucial for maritime safety, coastal management, and climate research. This thesis explores the application of advanced Machine Learning models including Linear Regression (LR), Convolutional Neural Networks (CN N ), Convolutional Neural Network-Long Short-Term Mem...
| Published: |
Swansea University, Wales, UK
2024
|
|---|---|
| Institution: | Swansea University |
| Degree level: | Doctoral |
| Degree name: | Ph.D |
| Supervisor: | Karunarathna, H. U., & Giannetti, C. |
| URI: | https://cronfa.swan.ac.uk/Record/cronfa70078 |
| Abstract: |
Accurate wave height prediction is crucial for maritime safety, coastal management, and climate research. This thesis explores the application of advanced Machine Learning models including Linear Regression (LR), Convolutional Neural Networks (CN N ), Convolutional Neural Network-Long Short-Term Memory (CN N -LST M ), Random Forest Regressor (RF R) and Extreme Gradient Boosting (XGBoost) to predict nearshore wave heights for Looe Bay and Felixstowe. For the purpose of site-specific offshore data extraction, differences in the complete-ness and quality of databases across seasons have been taken into account. A thorough coverage of different nearshore sites-from Isles of Scilly (2015-2019) and West Gabbard (2019-2022) is shown to account for different periods and conditions. A comprehensive search of 31 feature sets is conducted to optimise the model’s performance.Feature selection is implemented out to obtain the relevant features, and parameter tuning is applied to the model parameters, providing two distinct layers of optimisation. The models are compared against each other in a comparison experiment that analysed the predictive accuracy of using the model against different periods of data and for various regional wave dynamics.The results indicate that both deep learning and ensemble models can accurately predict wave heights in both locations despite data length and wave behaviour being different across the studies. The XGBoost model has been very successful in predicting both spatial and temporal wave structures, while CN N and CN N −LST M models still performed well, which emphasisestheir capabilities in portraying complex waves. Additionally, this study addresses challenges such as data dependency, overfitting and the computational demands of large-scale datasets. This thesis presents a framework that significantly enhances wave height forecasting for both locations by integrating features such as wave height (Hs), wave direction (Dir), wave period (T m), wave spread (Spr) and wave speed (Cv). The results highlight the adaptability and strength of the models, showing that both deep learning and ensemble methodologies can reach high levels of predictive power with careful optimisation of parameters, as well as detailed datasets. Furthermore, this study extends the use of Machine Learning in the area of oceanography, making a significant step towards more advanced systems of wave prediction in the future, with potential implications for coastal management focused on resilience. |
|---|---|
| Item Description: |
A selection of content is redacted or is partially redacted from this thesis to protect sensitive and personal information. |
| Keywords: |
Nearshore wave height, Machine Learning, XGBoost, coastal engineering, time-series forecasting, UK coast |
| College: |
Faculty of Science and Engineering |