Meandering compound channels are complex fluvial systems characterized by intricate hydraulic interactions between the primary channel and surrounding floodplains. These systems present significant challenges in flood risk management, sediment transport analysis, and ecological sustainability. Traditional hydrodynamic models, such as Saint-Venant equations and Computational Fluid Dynamics (CFD), often struggle with the computational cost and uncertainties associated with natural variability. In recent years, deep learning models, particularly Recurrent Neural Networks (RNNs), Long Short-Term Memory (LSTM), and Convolutional Neural Networks (CNNs), have emerged as powerful tools for modelling the dynamic behaviour of these channels. This study compares the performance of RNN, LSTM, and CNN in predicting discharge in meandering compound channels using hydraulic and geometric parameters such as bed slope, Manning’s roughness coefficient, and channel meandering characteristics. The dataset, sourced from observed or simulated values, was preprocessed and normalized to facilitate model training. The results demonstrate that while RNNs capture temporal dependencies, LSTMs excel in learning long-term dependencies crucial for riverine system modelling. However, CNNs, traditionally used for spatial data, outperformed both RNN and LSTM in predicting discharge, achieving the lowest Mean Squared Error (MSE) and highest R2 score. The CNN model’s ability to extract localized spatial patterns contributed to its robust performance, even in time series data. This study highlights the superior predictive accuracy and generalization capability of CNNs in meandering compound channel discharge prediction and discusses the potential for hybrid models that integrate temporal and spatial feature extraction. The findings also underscore the importance of incorporating deep learning techniques in hydrodynamic modelling, offering enhanced flood forecasting and management strategies.