Optimization of LSTM and GRU Deep Learning Models for Tidal Sea Level Time Series Prediction to Support Early Warning Systems

Abstract

Coastal regions are increasingly vulnerable to hazards associated with tidal sea level variability, including recurrent tidal flooding that threatens coastal infrastructure and communities. Accurate tidal prediction is therefore essential to support effective coastal early warning systems, particularly in regions with diverse tidal characteristics such as Indonesia. This study aims to optimize and compare Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) deep learning models for tidal sea level time series prediction. The analysis uses observational tidal data from four stations located in West Sumatra and North Sumatra, covering a continuous period from October to December 2025. The proposed framework includes data preprocessing, time series construction, baseline model development, and systematic hyperparameter tuning using Bayesian Optimization. Model performance is evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Percentage Error (MPE), and the coefficient of determination (R²). The results indicate that Bayesian Optimization significantly enhances the predictive accuracy of both models. LSTM achieves the best performance at Bungus, an open-ocean station with more complex tidal dynamics, while GRU outperforms LSTM at Belawan, Kuala Tanjung, and Sibolga, which exhibit more regular tidal patterns. These findings demonstrate that model performance depends strongly on local tidal characteristics rather than on model architecture alone. Overall, the study concludes that optimized LSTM and GRU models provide a reliable and adaptable approach for tidal sea level prediction and offer strong potential to support location-specific coastal early warning systems.