A Comparative Study of GRU and LSTM Time-Series Forecasting for Precise River Dam Hydrodynamic Prediction

Hydrodynamic forecasting plays a critical role in managing water resources and mitigating flood risks, particularly in urban areas. This study explores the application of machine learning and deep learning techniques to capture complex temporal patterns, focusing on water level, water inflow, and water outflow predictions. Utilizing the PMD-FFD dataset, we conducted empirical analyses to evaluate the performance of various models. Specifically, we evaluated 10 different model architectures, comprising 5 configurations each for LSTM and GRU, to identify the most effective setups for each parameter. Our study reveals insights, including the superiority of certain models for specific forecasting tasks. For instance, the GRU_2 model emerged as the most effective for water level forecasting with R
²
= 0.998, while LSTM_2 exhibited superior performance for water inflow forecasting with R
²
= 0.991. Moreover, our investigation into multi-step forecasting uncovered valuable findings. LSTM_5 demonstrated unexpected efficacy in multi-step predictions for most water parameters, with the highest MRE of 0.010, indicating its potential for capturing long-term temporal dependencies.