Deep learning-based forecasting of significant wave height in Padang Coastal Waters: a comparative study of RNN, LSTM, and GRU architectures

This study investigates deep learning–based forecasting of significant wave height (SWH) in Padang coastal waters, western Sumatra, a region exposed to wave-induced hazards that affect navigation, fisheries, and coastal infrastructure. SWH data are obtained from the ERA5 reanalysis single-levels product using the variable significant_height_of_combined_wind_waves_and_swell at hourly resolution from January to November 2025. After conversion from NetCDF to CSV, the time series is cleaned, normalized using Min-Max scaling, and transformed into supervised samples through a sliding-window approach. The dataset is split chronologically into training (January–September), validation (October), and testing (November) subsets. Three recurrent neural architectures are compared under a consistent experimental setup: a vanilla Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU). All models use the same input window length, optimization settings, and early-stopping criterion, and are evaluated with Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination . Results show that all architectures capture SWH dynamics with very high skill, yielding test values above 0.997 and MAPE below 0.7%. The RNN achieves the lowest test RMSE (0.007551) and MAE (0.003894), while the GRU delivers slightly higher errors but a modeling compromise between accuracy and training time. The LSTM attains the largest error among the three models (RMSE 0.008308, MAE 0.004423) yet is the most computationally efficient. Residual analyses indicate that the largest errors occur during sharp transitions and high-energy events. Overall, the study demonstrates that recurrent deep learning models driven solely by ERA5 reanalysis can provide accurate short-term SWH forecasts to support coastal and maritime decision-making in Padang and other data-sparse regions.