Ponentes
Descripción
Improving Paraguay River Water Level Predictions with Advanced Deep Learning Techniques
Giuliano Gonzalez, Nelson Ruiz, Diego H. Stalder, and Diego P. Pinto-Roa
Universidad Nacional de Asunción
San Lorenzo, Paraguay.
giuli1297.gg@fpuna.edu.py, nelsonruiz95@fpuna.edu.py, dstalder@ing.una.py, dpinto@pol.una.py
Accurate prediction of river water levels is crucial for effective flood and drought management, helping to protect communities and transportation infrastructure. Traditional hydrological models often face challenges due to seasonal variations influenced by rainfall, climate change, and land-use changes. This study addresses these challenges by employing deep learning techniques to forecast water levels in the Paraguay River at the port of Asunción, using a dataset with daily records since 1974 and additional stations at Rosario, Concepcion, and Fuerte Olimpo. A previous work identified the Gated Recurrent Unit (GRU) model as the most accurate baseline model for this task. The primary focus of this study is a time series cross-validation experiment with an expanding window, retrained annually, to determine how often the model should assimilate new data. The results indicate that this retraining strategy outperforms the baseline model, achieving lower Root Mean Squared Error (RMSE) values—0.5650 compared to 0.6088 and lower Nash–Sutcliffe model efficiency coefficient (NSE) values—0.9327 compared to 0.9177. This highlights the importance of strategic retraining alongside initial model parameters to enhance accuracy over time. These findings pave the way for refining hydrological predictions, supporting sustainable water resource management, and enhancing preparedness for extreme weather events. This study provides valuable tools for informed decision-making in water management and natural disaster risk reduction for vulnerable communities.
Keywords: River Water Level Predictions, Deep Learning
Improving Paraguay River Water Level Predictions with Advanced Deep Learning Techniques
