StormNet: GNN model improves storm surge predictions
A new spatio-temporal graph neural network, named StormNet, has been created to enhance the accuracy of storm surge predictions. This innovative model combines long short-term memory (LSTM) elements with graph convolutional (GCN) and graph attention (GAT) techniques to effectively capture intricate spatial and temporal relationships among water-level monitoring stations. Utilizing historical hurricane data from the U.S. Gulf Coast and tested on Hurricane Idalia (2023), StormNet significantly lowers root mean square error when compared to conventional high-fidelity numerical models such as ADCIRC. The research tackles the uncertainties present in traditional models, especially in light of recent patterns of rapid intensification and heightened nearshore storm occurrences.
Key facts
- StormNet is a spatio-temporal graph neural network for storm surge bias correction.
- It integrates GCN, GAT, and LSTM components.
- Trained on historical hurricane data from the U.S. Gulf Coast.
- Evaluated on Hurricane Idalia (2023).
- Reduces root mean square error compared to ADCIRC.
- Addresses uncertainties in traditional numerical models.
- Focuses on rapid intensification and increasing nearshore storm activity.
- Published on arXiv with ID 2604.20688.
Entities
Institutions
- arXiv
Locations
- U.S. Gulf Coast