Sustainable water management is a major challenge in semi-arid regions, where climate variability and irregular rainfall intensify pressure on reservoir systems. In this context, forecasting reservoir water levels becomes essential for optimizing water allocation, anticipating shortages, and strengthening hydrological resilience. The Ahmed El Hansali Dam, located in the Oum Er-Rbia basin, plays a strategic role in domestic water supply, irrigation, and hydropower generation. However, traditional forecasting methods often struggle to capture the nonlinear and complex dynamics of hydrological time series.
The rise of machine learning, particularly Recurrent Neural Networks (RNNs), provides new opportunities for modeling such complex behaviors. This work evaluates three neural architectures — ANN, Simple RNN, and GRU — to predict monthly reservoir levels using nearly 50 years of hydro-climatic data.
The study was conducted in collaboration with the Directorate General of Hydraulics (DGH), Ministry of Equipment and Water of Morocco. Its objective is to identify the most effective model to support data-driven decision-making for reservoir management.
Beyond its operational value, this research contributes to understanding the relevance of deep learning architectures for hydrological forecasting in semi-arid environments.