<p>Simulating daily discharge with Deep Learning (DL) models poses a unique challenge due to periods of seasonal zero flow. This study proposes a framework for implementing DL techniques to simulate daily river discharge, particularly in rivers with seasonal flow variations including periods of zero flow. Periods of seasonal no-flow result in a long sequence of zeros in the target data. This sparse data representation is a challenge in DL models applied for daily discharge simulations, as it can lead to physically unrealistic negative flow predictions. The selection of appropriate data preprocessing techniques, weight initialization, activation functions, regularization, and adaptive learning techniques are crucial while dealing with sparse data representation. The framework so developed is demonstrated in three popular deep learning techniques- Recurrent Neural Network (RNN), Long Short Term Memory (LSTM), and Gated Recurrent Units (GRU) for discharge simulations. These models simulate daily discharge using the meteorological inputs- precipitation, maximum and minimum temperature, and relative humidity. All the models were optimised without yielding any negative flow predictions. It was found that both LSTM and GRU performed equally well and better than RNN in the entire range of flow. In no-flow conditions, LSTM remarkably outperformed the other two, and GRU was superior to RNN. Thus, reframing the deep learning models will enhance the overall reliability, particularly in drought studies and water allocation policies.</p>