Dang Truong An, Tran Thuy Tien, Tran Binh Nguyen, Nguyen Cong Thanh and Pham Bao Quoc
Adv. Artif. Intell. Mach. Learn., 5 (4):4418-4432
1. Dang Truong An: Department of Oceanology, Meteorology and Hydrology,University of Science, HCM City 749000, Vietnam; VNU-HCM, HCM City 700000,Vietnam.
2. Tran Thuy Tien: Department of Oceanology, Meteorology and Hydrology, University of Science
3. Tran Binh Nguyen: Department of Oceanology, Meteorology and Hydrology, University of Science, HCM City 749000
4. Nguyen Cong Thanh: Department of Oceanology, Meteorology and Hydrology, University of Science, HCM City 749000
5. Pham Bao Quoc: Faculty of Natural Sciences, Institute of Earth Sciences, University of Silesia in Katowice, Będzińska street 60, 41-200, Sosnowiec, Poland
DOI: 10.54364/AAIML.2025.54245
Article History: Received on: 16-Jul-25, Accepted on: 16-Oct-25, Published on: 23-Oct-25
Corresponding Author: Dang Truong An
Email: dtan@hcmus.edu.vn
Citation: Tran Thuy Tien, Tran Binh Nguyen, Nguyen Cong Thanh, Pham Quoc Bao, Dang Truong An. Forecasting Water Level of the Vietnamese Mekong Delta Integrating the Harmonic Tidal Method and Deep Learning Model. Advances in Artificial Intelligence and Machine Learning. 2025 .
Hourly water level forecasting (HWLF) is essential for flood
management and disaster mitigation in hydrodynamically complex regions like the
Vietnamese Mekong Delta (VMD). Here, we benchmarked a Long Short-Term Memory
(LSTM) network for HWLF at the My Thuan station on the Tien River, a location
influenced by both river and tidal dynamics. Using an hourly dataset from 1978
to 2022, we compared the LSTM model's predictions against the traditional
harmonic analysis method for lead times of 1 to 168 hours. A sensitivity
analysis showed that an input sequence of 360 past hourly observations (15
days) was optimal. The LSTM model performed well at short lead times, with a
NASH index of 0.98 and an RMSE of 9.25 cm for a 1-hour forecast. However, its
accuracy decreased significantly at longer horizons; the NSE dropped to 0.59
and the RMSE exceeded 48 cm at 168 hours. Despite this, the LSTM model was
consistently better than the harmonic method, especially at capturing the
non-linear interactions between river discharge and tides. Our results provide
a clear performance baseline for a univariate deep learning model in this
region, defining its operational limits. While LSTM is promising for short-term
forecasts, our findings show that external physical data is needed to improve
long-term reliability.