Wirthensohn et al. (2026) Learning water diversion behaviour in regulated river networks using graph neural networks and neural decision trees

Identification

• Journal: Journal of Hydrology
• Year: 2026
• Date: 2026-09-03
• Authors: Moritz Wirthensohn, Ye Tuo, Markus Disse
• DOI: 10.1016/j.jhydrol.2026.136350

Research Groups

• Chair of Hydrology and River Basin Management, Technical University of Munich

Short Summary

This study investigates how graph-based deep learning models can represent water diversion behaviour in regulated river networks under varying degrees of prior operational knowledge. The results show that different levels of operational knowledge affect the ability of graph-based models to represent water diversion behaviour.

Objective

• Quantify how different degrees of operational knowledge affect the ability of graph-based models to represent water diversion behaviour in regulated river networks.
• Assess the implications of jointly learning hydrological dynamics and operational decisions versus representing them as explicitly separable components.
• Analyse how prediction errors propagate between hydrological states and operational decisions under alternative graph-based model strategies.

Study Configuration

• Spatial Scale: A regulated alpine subcatchment of the Upper Isar River basin in southern Germany, with a spatial resolution of 30 m.
• Temporal Scale: Daily time step, covering the period 1983–2020.

Methodology and Data

• Models used: Graph Attention Network (GAT) architecture, including four model configurations: GATBase, GATEdge, GATPost, and GATNDT.
• Data sources: SWAT+ simulations as an internally consistent synthetic benchmark, with meteorological forcing data from the SDCLIREF dataset.

Main Results

• Graph-based models can reproduce total flow before diversion robustly, but diverted flow skill declines with operational complexity.
• Additional operational knowledge does not consistently improve predictive accuracy, as diversion errors are dominated by propagation from total-flow uncertainty.
• The hybrid model (GAT_NDT) achieves comparable skill while providing explicit and analysable decision representations.

Contributions

• This study contributes to the development of ML-based hydrological models that explicitly account for human decision-making in regulated river systems.
• The results highlight design choices for future applications where operational data are incomplete or uncertain, and provide insights into how different graph-based architectures represent managed processes in river networks.

Funding

• Not specified

Citation

@article{Wirthensohn2026Learning,
  author = {Wirthensohn, Moritz and Tuo, Ye and Disse, Markus},
  title = {Learning water diversion behaviour in regulated river networks using graph neural networks and neural decision trees},
  journal = {Journal of Hydrology},
  year = {2026},
  doi = {10.1016/j.jhydrol.2026.136350},
  url = {https://doi.org/10.1016/j.jhydrol.2026.136350}
}