Wasserstein temporal GAN for ocean wave prediction with small training sets

Wave prediction is becoming increasingly important to provide optimal solutions in marine activities such as autonomous decision making and human activity planning. Learning-based algorithms require a sufficient amount of historical data for training, however, in marine scenarios, the acquisition of valid, high-quality wave data is frequently constrained by factors such as equipment limitations, harsh ocean conditions, or limited monitoring coverage. To address the challenge of sea wave prediction with small-scale training datasets, this paper proposes a new data augmentation model, WTGAN (Wasserstein Temporal Generative Adversarial Network). By capturing temporal dependencies and generating synthetic time series data, the WTGAN model aims to mitigate the data scarcity issue. The effectiveness of the new proposed model is validated by comparing the statistical properties of synthetic data, as well as the performance of three commonly-applied forecasting models with and without data augmentation. The performance of sea wave forecasting is assessed by the Mean Absolute Error (MAE), Root Mean Square Error (RMSE) and Coefficient of Determination ( R 2 ), and is improved with the proposed WTGAN model by 5% to 12% under different wave conditions.