Forecasting Water level at Khanh An station, An Giang province using LSTM deep learning model

Abstract

    Water level forecasting is a vital tool in water resource management, especially for regions affected by flooding during the rainy season and water scarcity during the dry season, such as the Mekong Delta area. In this area, water level fluctuations exhibit a seasonal cycle, influenced simultaneously by tidal regimes and upstream flows. As such, forecasting methods must be capable of handling nonlinear time series data and outperform traditional approaches. This study employs a deep learning model, specifically the Long Short-Term Memory (LSTM) model, to forecast water levels at Khanh An station, situated in An Giang Province, a key monitoring gauge in the Hau River hydrological system. Two forecasting scenarios were developed, using 24-hour and 48-hour input sequences to predict the next 6 hours of water levels. The model was trained with number of epochs (50, 100, 200, and 300). Results showed that the model performs best with 48-hour input data and 300 epochs, achieving a Root Mean Square Error (RMSE) of 6.894 and a coefficient of determination (R²) of 0.997 on the test set. The model accurately simulates extreme conditions and serves as an effective tool for seasonal water level forecasting at Khanh An station. It holds strong potential for broader application in flood warning, and drought management in regions significantly impacted by climate change and flow variability, such as the Mekong Delta in general and An Giang Province in particular.