Session: 06-17-02 AI Technology for Ocean Engineering

Submission Number: 156549

A Methodological Approach to Short-Term Coastal Wave Prediction Using Combined Pre-Analysis and Machine Learning 

The practical application of clean energy is a key research topic today. Coastal wave energy conversion, transforming wave energy into electrical energy, is a potent method for wave energy development. Site selection is crucial in the development process, requiring an integrated assessment of wave resources potential, safety, and economic viability. Additionally, precise prediction technology plays a significant role in guiding site selection and optimizing resource allocation. The historical data from numerous observation stations along Japan’s coast make predicting short-term wave conditions relatively feasible. For unobserved sea areas, estimating wave characteristics using data from adjacent regions not only fills data gaps but also enhances the scientific depth and practical applicability of the research.

Contemporary models for coastal wave prediction, such as the SWAN model, require extensive oceanographic datasets that include parameters such as wave height, period, temperature, atmospheric pressure and more. Initiating comprehensive data collection efforts at this juncture would substantially escalate research expenditures. Furthermore, while machine learning techniques have seen widespread application in forecasting wave conditions in observed locales, explorations into unmonitored regions remain limited. Notably, the precision of predictions for wave conditions tends to wane with extended forecast horizons.

The NOWPHAS network represents Japan's extensive coastal wave monitoring infrastructure, encompassing over 80 locations nationwide with continuous wave observation. With inter-site distances ranging from 15 to 40 kilometers, the extensive long-term data accrued are pivotal in coastal development, disaster prevention, and related domains. This investigation employs these data to formulate a methodology capable of predicting wave conditions, including wave height and period, up to 24 hours in advance at both monitored and unmonitored sites.

This research initially undertakes a thorough analysis of wave characteristics across diverse water depths and seasonal variations. Subsequent evaluations of the time delays associated with wave propagation allow for an analysis of wave transmission modes under varying conditions. By employing machine learning algorithms, specifically Long Short-Term Memory (LSTM) networks, this study aims to predict wave conditions across as many maritime regions as feasible along all coastal fronts. This approach aids in identifying wave distribution patterns, investigating principal factors influencing wave propagation, and establishing a maximum prediction distance of approximately 20 kilometers from observed sites, thus ensuring the reliability of forecasts within a 24-hour timeframe.

By delving into the intricacies of numerical modeling, this methodology is crafted to minimize observational expenses and maximize both the extent and duration of predictive capabilities. Particularly, the development of this methodology offers profound insights into optimizing resource allocation, enhancing model precision, and broadening the operational applicability of wave prediction frameworks.

Presenting Author: Sijia Wang The University of Tokyo

Presenting Author Biography: Wang Sijia is a second-year doctoral student at the Graduate School of Frontier Sciences at The University of Tokyo, specializing in Ocean Technology, Policy, and Environment. She is a member of the Rheem Laboratory, where her research focuses on the observation and prediction of sea surface conditions.