This review critically examines the use of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Gated Recurrent Units (GRUs) in hydrological time series predictions. RNNs, while foundational, face limitations such as vanishing gradients, which hinder their ability to model long-term dependencies. LSTMs and GRUs have been developed to overcome these issues, with LSTMs using memory cells and gating mechanisms, and GRUs providing a more streamlined architecture. The integration of attention mechanisms and hybrid models that combine RNNs, LSTMs, and GRUs with other Machine Learning (ML) and Deep Learning (DL) techniques has improved prediction accuracy by capturing both temporal and spatial dependencies. Despite their effectiveness, practical implementations require extensive datasets and substantial computational resources. Future research should focus on developing interpretable architectures, enhancing data quality, incorporating domain knowledge, and utilizing transfer learning to improve model generalization and scalability across diverse hydrological contexts. The review covers the theoretical foundations of RNNs, LSTMs, and GRUs, their architectural differences, and their applications in hydrological time series prediction. It highlights the advantages and limitations of each model, particularly in handling long-term dependencies, noisy data, and missing values. The effectiveness of hybrid models, such as LSTM-CNN and LSTM-GRU, is also discussed, along with their impact on prediction accuracy and computational efficiency. The review concludes by addressing key research questions, including the architectural differences between RNNs, LSTMs, and GRUs, the adaptations and modifications to improve LSTM performance, the implementation and effectiveness of these models in real-world hydrological applications, and the trends in the application of hybrid models over the past decade.This review critically examines the use of Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTMs), and Gated Recurrent Units (GRUs) in hydrological time series predictions. RNNs, while foundational, face limitations such as vanishing gradients, which hinder their ability to model long-term dependencies. LSTMs and GRUs have been developed to overcome these issues, with LSTMs using memory cells and gating mechanisms, and GRUs providing a more streamlined architecture. The integration of attention mechanisms and hybrid models that combine RNNs, LSTMs, and GRUs with other Machine Learning (ML) and Deep Learning (DL) techniques has improved prediction accuracy by capturing both temporal and spatial dependencies. Despite their effectiveness, practical implementations require extensive datasets and substantial computational resources. Future research should focus on developing interpretable architectures, enhancing data quality, incorporating domain knowledge, and utilizing transfer learning to improve model generalization and scalability across diverse hydrological contexts. The review covers the theoretical foundations of RNNs, LSTMs, and GRUs, their architectural differences, and their applications in hydrological time series prediction. It highlights the advantages and limitations of each model, particularly in handling long-term dependencies, noisy data, and missing values. The effectiveness of hybrid models, such as LSTM-CNN and LSTM-GRU, is also discussed, along with their impact on prediction accuracy and computational efficiency. The review concludes by addressing key research questions, including the architectural differences between RNNs, LSTMs, and GRUs, the adaptations and modifications to improve LSTM performance, the implementation and effectiveness of these models in real-world hydrological applications, and the trends in the application of hybrid models over the past decade.