A federated learning based approach for wave height prediction in maritime transportation

Forecasting Ocean Wave Height (OWH) is essential for several marine operations, such as coastal and ocean engineering applications. Current state-of-the-art solutions in the domain involve building machine and deep learning-based models for estimating wave height patterns. These existing solutions lack the desired generalization capability and heavily rely on data, contextual and external features. To address these limitations, a federated learning-based approach utilizing long short-term memory (LSTM) network and Temporal Convolutional Network (TCN) for ocean wave height estimation is proposed in the current work. The experimental evaluation of the proposed solution is performed in a federated learning setting, utilizing an open-source dataset comprising eight selected National Data Buoy Center (NDBC) buoys placed in various maritime regions. The aim is to demonstrate how comparable or even better performance than individual learning may be achieved by training the individual models on decentralized time-series data and then aggregating their updates to the global model. Through experimental evaluation, the resultant prediction accuracy, generalization, and computational benefits of the proposed approach are verified compared to the existing benchmark studies.