Exploring Neural Network Performance in Hydrological Modeling in a Mountainous Region of Morocco: A Case Study on LSTM and GRU Architectures for Runoff Prediction

This study thoroughly compared two types of neural networks, namely Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) recurrent neural networks (RNNs), within the context of hydrological modeling for predicting runoff in the Oum Er Rabia sub-basins. By assessing their performance on both daily and monthly scales, we aimed to understand the dynamics of runoff, which is crucial in water resources management. The combined analysis of predictions at both time scales provides a comprehensive perspective, using daily outputs for short-term decisions and monthly predictions for longer-term planning.

Using a hydroclimatic time series dataset from 2000 to 2019, incorporating key factors such as snow cover area, temperature, and rainfall that influence hydrological processes and significantly impact flow patterns, the research evaluates the predictive accuracy of the models at both scales. The results reveal nuanced differences in predictive accuracy, with average Kling-Gupta Efficiency (KGE) values for LSTM and GRU at daily and monthly scales, respectively, being 0.64, 0.52, and 0.46, 0.54. These findings provide insights into the strengths and limitations of each architecture in the mountainous region of Morocco.

The study enhances our understanding of the applicability of LSTM and GRU architectures in hydrological modeling, aiding practitioners in selecting models tailored to specific needs. By establishing a robust framework for short-term decision-making and long-term planning in water resource management, this research contributes to advancing predictive modeling and promoting sustainable water use while mitigating flood risks. The knowledge acquired paves the way for improved decision support in the critical area of water resource management.