A physics-informed hypergraph attention network for forecasting streamflow drought index
PHYSICS AND CHEMISTRY OF THE EARTH, cilt.145, 2027 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 145
- Basım Tarihi: 2027
- Doi Numarası: 10.1016/j.pce.2026.104889
- Dergi Adı: PHYSICS AND CHEMISTRY OF THE EARTH
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Chimica, Compendex, Geobase, INSPEC
- Ondokuz Mayıs Üniversitesi Adresli: Evet
Özet
This study proposes a novel Physics-Informed Temporal-Spatial Hypergraph Attention Network (PI-TSHAN) model to predict the Streamflow Drought Index (SDI) using data from seven streamflow gauging stations within the K & imath;z & imath;l & imath;rmak Basin, T & uuml;rkiye. The PI-TSHAN architecture goes beyond standard graph neural networks (GNNs), utilizing hypergraph convolutions that model multi-station interactions within a watershed as high-order relationships. It analyzes spatial dependencies using spectral hypergraph theory while learning temporal dynamics through multi-scale attention mechanisms. Furthermore, differential hydrological constraints and mass conservation principles included in the model's loss function ensure that predictions are not only data-driven but also physically consistent. The experimental evaluation was conducted comparatively with state-of-the-art (SOTA) algorithms such as Informer, Autoformer, PatchTST, and ST-iTransformer at SDI-1, SDI-3, and SDI-12 scales. The findings showed that the PI-TSHAN model achieved the highest overall accuracy, ranking first at the majority of the stations and accumulation scales and remaining competitive with the strongest baselines elsewhere. In particular, in long-term predictions (SDI-12), the model demonstrated near-perfect fit, achieving NSE of 0.9944 and KGE of 0.9811 at station D15A152. In short-term and highly volatile SDI-1 data, the model achieved Drought Duration Accuracy (DDA) exceeding 97%, resulting in 15-20% higher accuracy in detecting extreme events compared to traditional models. The numerical results demonstrate that integrating physical constraints consistently improves the generalizability and prediction stability of the model, especially in data-constrained regions. In conclusion, PI-TSHAN offers a reliable early warning tool for the K & imath;z & imath;l & imath;rmak Basin and similar complex catchments, though broader generalizability across T & uuml;rkiye warrants future research.