Water reservoirs are essential for supply, flood control, and hydropower, but their operations can disrupt natural river flows and complicate water management, planning, and fair resource allocation. As a result, there is growing interest in data-driven models capable of predicting complex reservoir behavior using historical data. Among these, deep learning algorithms, particularly recurrent neural networks (RNNs) such as Gated Recurrent Units (GRUs) and Long Short-Term Memory (LSTM) networks, have shown strong potential for reservoir operation modeling and hydropower prediction. This study applies GRU and LSTM models to the Hoover Dam in the Colorado River Basin, using historical data from 1938 to 2023. A systematic hyperparameter search was conducted, with 2,500 training runs per model, exploring various configurations of hidden layers, hidden units, and batch sizes. The results show that single-layer models offer the best balance between accuracy and computational efficiency for this dataset. However, for applications demanding greater scenario robustness, two-layer architectures provide improved accuracy and enhanced representational capacity with only a modest increase in computational overhead. GRU models performed best with fewer than 125 units and batch sizes under 150, while LSTM models were most effective with approximately 100 units and batch sizes below 50. Increasing the number of layers led to longer training times without accuracy gains. Both models achieved strong validation results (NSE > 0.75, RSR < 0.5), effectively modeling daily discharge. These findings highlight the potential of GRU and LSTM networks for reliable, efficient reservoir forecasting under complex and changing conditions.