Publication

Abstract : The hydrological community is focusing on streamflow forecasting due to rising water consumption and climate change impacts. Traditional AI techniques help understand surface water hydrology, but consistent performance remains challenging due to insufficient feature learning. To address these challenges, this study introduces a novel multi-tier deep learning architecture based on the Visual Geometry Group 19 (VGG19) baseline model. The proposed architecture, named MTV19ANet, integrates a multi-attention-based fused framework combining Long Short-Term Memory (LSTM), Multi-Layer Perceptron (MLP) neural network, and Gated Recurrent Unit (GRU). MTV19ANet modifies the original VGG19 architecture by reducing convolutional blocks and decreasing computational load while preserving core feature extraction functionality. Attention mechanisms are used to weigh important features, while Global Average Pooling (GAP) reduces spatial dimensions, summarizing features for the attention mechanism. The study addresses key challenges in hydrological modeling, such as capturing complex spatiotemporal patterns and managing computational efficiency. The model was tested using the CAMELS dataset, covering catchments in Australia, Brazil, and Great Britain. Performance comparisons with VGG19, LSTM, GRU, and MLP demonstrated that MTV19ANet consistently outperforms these models, achieving Nash–Sutcliffe Efficiency (NSE) values of 0.986, 0.991, and 0.985 for the CAMELSAUS, CAMELSGB, and CAMELSBR datasets, respectively. The results indicate that this approach can be applied to other catchments, with different climatic and geographical conditions, and increase the accuracy of streamflow predictions for other parts of the world. This study provides a replicable framework that incorporates physical hydrological attributes with dynamic meteorological data for improved predictive accuracy.