Floods are among the most catastrophic natural disasters, particularly in India, where diverse river systems and monsoonal rain pose significant risks. Traditional hydrological models, while helpful, often fail to capture complex, nonlinear interactions within watersheds under various meteorological conditions. This study introduces advanced Deep-Learning (DL) models, specifically Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU), enhanced with an attention mechanism to improve rainfall-runoff modelling for flood forecasting in the Krishna River. Our results show that these models significantly outperform traditional DL methods such as LSTM and GRU, with improvements in the Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R \(^{2}\) ). LSTM-GRU, with an attention hybrid model, demonstrated superior performance, highlighting its ability to focus on critical parts of the input sequence and provide more accurate predictions. Despite these advances, the dependency of models on large datasets, their computational intensity, and issues with interpretability are challenges that need to be addressed.