IVI at NeurIPS 2025: Innovative AI Model for Hydrological Forecasting within the ARTIFACT Project

Researchers from the Institute for Artificial Intelligence of Serbia presented their work at one of the world’s most prestigious artificial intelligence conferences — NeurIPS 2025 (Conference on Neural Information Processing Systems).

Their study, titled “TC-GTN: Temporal Convolution Graph Transformer Network for Hydrological Forecasting”, was developed as part of the Horizon Europe TWINNING project ARTIFACT (Artificial Intelligence for Flood-Resilient Infrastructure). The research focuses on applying advanced AI methods for river flow forecasting, which is crucial for flood risk management, responding to extreme climate events, and the sustainable management of hydropower systems.

The authors of the study are IVI researchers Ana Samac, Dr. Milan Dotlić, Luka Vinokić, Dr. Milan Stojković, and Dr. Veljko Prodanović. In the context of increasingly severe climate change and rising flood frequency, accurate and reliable hydrological forecasting remains a key challenge in modern water resources management. Within the ARTIFACT project, the focus is on developing AI-supported solutions for flood-resilient infrastructure, and this research contributes directly to that goal.

The researchers highlight that most existing models do not fully leverage the directed and hierarchical graph structure of hydrological systems, such as river networks with their associated hydrological and meteorological stations. The proposed TC-GTN (Temporal Convolution Graph Transformer Network) represents a hybrid approach that combines:

• Temporal Convolutions (TC) for efficient extraction of time-dependent patterns;
• Graph Transformers (GT) for advanced modeling of spatial and structural dependencies within river systems.

By using an attention mechanism, the model provides a deeper understanding of interactions between different stations, while more accurately capturing time dependencies at individual locations.

The model was tested on real data from the Drina–Lim river basin. Experimental results show that TC-GTN outperforms existing baseline models in predicting regular flows as well as extreme high-water events, demonstrating its effectiveness in forecasting critical hydrological phenomena.

These results are particularly valuable for:

• Enhancing flood early warning systems;
• Improving risk management in urban and rural areas;
• Supporting sustainable planning and operation of hydropower systems under changing climate conditions.

This work is part of the international Climate Change AI initiative and contributes to increasing the global visibility of the ARTIFACT project. By participating in NeurIPS 2025, the Institute for Artificial Intelligence of Serbia further solidifies its position as a key international player in AI for climate resilience, highlighting the ARTIFACT project’s role in bridging cutting-edge research with tangible societal and infrastructure needs.