Application of the LSTM (Long Short-Term Memory) model for tidal forecasting: A case study at the Ba Ria -Vung Tau tide gauge station, Vietnam

Abstract

In recent years, the application of tidal research has markedly supported and enhanced various socio-economic activities in coastal regions, particularly port operations, aquaculture, and disaster prevention and mitigation. Accurate tidal water level forecasting plays a vital role in spatial planning, coastal infrastructure management, and early flood warning systems. This study proposes the application of a Long Short-Term Memory (LSTM) recurrent neural network model to forecast daily tidal water levels atthe Ba Ria -Vung Tau station, based on observed data from 1999 to 2022. The input data were preprocessed through three main steps: outlier removal using the Interquartile Range (IQR) method, Min-Max normalization, and time series decomposition into trend, seasonal, and residual components. The model was trained on the 1999÷2021 dataset, tested with 2022 data, and used to forecast tidal levels for 2023 using recursive forecasting. The results show that the LSTM model achieved high performance with low error on the test dataset (MSE = 0.0039 cm, MAE = 0.0449 cm, R² = 0.9443). The model effectively captured the semi-diurnal tidal cycle, demonstrated fast convergence, low error, and high stability, highlighting its potential for integration into automated tidal forecasting and early warning systems for coastal management. However, the current model uses only univariate data (tidal water level) without incorporating meteorological and oceanographic variables such as wind, pressure, temperature, or wave data. Future studies should expand toward multivariate models to improve generalization and forecasting accurac