A Spatiotemporal Machine Learning Framework for Nearshore Wave Modelling

Abstract

Accurate, real-time wave forecasting in the spatiotemporal domain is crucial for the sustainable operation of nearshore marine renewable energy devices. Machine Learning (ML) has shown great promise in nearshore wave modelling, reducing computational costs through its data-driven approach. Nevertheless, ML-based models face significant challenges. These include accurately capturing high-resolution nearshore waves in both spatial and temporal domains, dealing with missing values in observational data, and generalising the model for various conditions. To address these challenges, this thesis proposed a model framework that decomposed the spatiotemporal problem into discrete spatial and temporal components, each capitalising on the strengths of physics-based numerical model outputs and in-situ observations, respectively. This bifurcated approach enhances modelling efficiency. The spatial dimension is represented by a surrogate model that utilises a Multi-output Random Forest (MRF) model to extrapolate from discrete grid points with existing observations to the entire gridded domain. Meanwhile, the temporal dimension is captured through a Long Short-Term Memory (LSTM) neural network model, offering a 12-hour sequential forecast with a half-hourly interval for multiple buoy locations. This LSTM-based model benefits from a tensor completion approach that fills gaps within historical observation data. Integrating these spatial and temporal models promotes a comprehensive and effective spatiotemporal forecast model framework. Testing the model framework in the waters off Cornwall, UK, the proposed model matches the accuracy of advanced UK Met Office numerical forecasts in estimating significant wave height (Hs) and surpasses it in estimating zero-crossing wave period (Tz) at validation buoy locations for up to 12-hour forecast horizons, while requiring significantly less computational resources. The proposed model’s modular design provides flexibility, facilitating the replacement of spatial and temporal components, and its adaptability to diverse data sources allows the integration of ship-as-a-wave-buoy into the observing network. In conclusion, this model framework offers a low-cost, resource-efficient approach for delivering short-term forecasts and demonstrating compatibility with existing methodologies. Its performance, comparable to leading physics-based numerical wave predictions, indicates substantial potential in pioneering a new class of rapidly updating met-ocean capabilities.