Temporal Graph Attention Networks for Physics-Consistent Flood Wave Propagation Across River Cross-Section Networks

Pardhasaradhi Menda — B.Tech Computer Science and Engineering (AI/ML), SRM Institute of Science and Technology.

Research question

Can a temporal graph attention network forecast river discharge at cross-sections with no gauge, and do its attention weights recover each reach’s hydraulic travel time?

What was found

Accuracy, qualified. Attention (GAT) lowered ungauged discharge error by +7.79% against a GCN, pooled over eight synthetic basins (95% interval +2.5% to +13.6%). On the two held-out basins alone, +8.86% (95% interval −2.9% to +17.2%), qualified: transfer is not established.
Interpretability, did not hold. Attention does not track travel time: hop-stratified r = +0.017 on all message edges and −0.129 on routing edges, at chance. Reach length alone scores +0.438, higher than the learned attention, and attention agrees across seeds at only +0.386.
Real gauges, pending. Results on the real network’s four gauges (H1–H4) do not exist yet. Every number above comes from synthetic basins.

The interactive figures need JavaScript. The run outputs behind every number are in the repository.