This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Accelerating HEC-RAS: A Recurrent Neural Operator for Rapid River Forecasting
Created by
Haebom
Author
Edward Holmberg, Pujan Pokhrel, Maximilian Zoch, Elias Ioup, Ken Pathak, Steven Sloan, Kendall Niles, Jay Ratcliff, Maik Flanagin, Christian Guetl, Julian Simeonov, Mahdi Abdelguerfi
Outline
This paper proposes a deep learning-based surrogate model to speed up computationally expensive physics-based solvers such as HEC-RAS for rapid decision-making during flood events. Using HEC-RAS as a data generation engine, we use a hybrid autoregressive architecture that combines GRU to capture short-term temporal dynamics and Geo-FNO to model long-range spatial dependencies along stream segments. Trained on 67 segments in the Mississippi River basin, the model is evaluated on 1 year of unknown data and achieves high prediction accuracy with a median absolute water level error of 0.31 feet. In an ensemble prediction for all 67 segments, the model is approximately 3.5x faster, from 139 minutes for the conventional solver to 40 minutes. This data-driven approach demonstrates that powerful feature engineering can be used to generate viable and fast surrogate models that can replace conventional hydraulic models, thereby improving the computational feasibility of large-scale ensemble flood prediction.
Takeaways, Limitations
•
Takeaways:
◦
We present a deep learning-based alternative model that can dramatically reduce the computational cost of physics-based flood prediction models such as HEC-RAS.
◦
Effectively modeling temporal and spatial dependencies through a hybrid architecture combining GRU and Geo-FNO.
◦
Presenting the potential for improving real-time and efficiency of ensemble flood forecasting in large watersheds such as the Mississippi River Basin.
◦
Demonstrate the feasibility of a data-driven approach that can effectively replace traditional repair models through robust feature engineering.
•
Limitations:
◦
The performance of a model can be highly dependent on the quality and quantity of training data.
◦
Since this model is specific to the Mississippi River basin, further study is needed to determine generalizability to other basins.
◦
Because it learns from data without explicitly modeling physical processes, the physical interpretation of the predicted results may be difficult.
◦
Further validation of the accuracy and stability of long-term predictions is needed.
