TC-GTN: Temporal Convolution Graph Transformer Network for Hydrological Forecasting

Machine learning enables accurate streamflow forecasting, vital for managing increasingly frequent flood events under climate change. However, most existing approaches do not fully exploit the inherent directional and hierarchical graph structure of hydrological systems. This paper introduces TC-GTN (Temporal Convolution Graph Transformer Network), a hybrid model designed for streamflow forecasting that integrates temporal convolution (TC) with graph transformers (GT). It uses the combination of TC for temporal pattern extraction and GT for advanced relational reasoning. It utilizes a structured graph representation of the river network with accompanying meteorological stations where the transformer's attention mechanism is critical for a better understanding of interactions between different nodes/stations and for capturing self-dependencies within each station.Experiments on the Drina–Lim River Basin dataset show that TC-GTN model outperforms baseline methods for regular flow rates, and also demonstrate improvements for high flow rates, which represent extreme hydrological events. Such performance is critical for effective flood risk mitigation and sustainable hydropower management under climate change effects. Code is available at: https://github.com/dodi007/TC-GTN-Spatio-temporal-Graph-Transormer.git.