Fully Distributed Rainfall-Runoff Modeling Using Spatial-Temporal Graph Neural Network
Abstract
Recent studies using the latest deep learning algorithms such as LSTM (Long Short-Term Memory) have shown great promise on time-series modeling. These studies focused on the watershed-scale rainfall-runoff modeling or the streamflow forecast, but most of them are limited to only considering a single watershed as a unit. Although this simplification is very effective, it ignored the spatial information which may cause significant errors in large watersheds. Several studies explored the use of data integration with GNN (Graph Neural Networks) by decomposing a large watershed into multiple sub-watersheds, but each sub-watershed is still considered as a whole and the geoinformation inside the watershed is not fully utilized. In this study, we proposed a novel deep learning model GNRRM (Graph Neural Rainfall-Runoff Model), which utilizes GNN to make full use of spatial information including the flow direction and geographic information from the high-resolution precipitation data. Specifically, we applied a time-series model on each grid cell for its runoff production. The output of each grid is then aggregated by a GNN as the final runoff at the watershed outlet. Our case study shows that our GNN module works successfully for high-resolution precipitation data. GNRRM has shown less over-fitting and has a significant improvement on the model performance compared to the baseline models. Our research further confirms the importance of hydrological information in rainfall-runoff modeling using deep learning, and we encourage researchers to incorporate more domain knowledge in modeling.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2021
- Bibcode:
- 2021AGUFM.H33J..06X