Learning the streamflow spatial and temporal dynamics through AI/ML
Abstract
Process-based hydrological models are broadly used to predict streamflows, but running hyper-resolution river models over large river basins is still computationally demanding. Many data-driven and hybrid ML models have been introduced for streamflow forecasting. A number of those studies focused on the large-sample hydrology datasets, in which AI/ML models are trained to learn a latent representation of inter-basin dynamics by leveraging the observed basin attributes from hundreds of basins. At the basin scale, additional complexities arise because of the non-Euclidean geometry of river networks and the accompanied non-uniformity in spatial and temporal flow dynamics. Connecting these disparate intra-basin scales is important for developing high-fidelity digital twins of river basin models. In this work, we adapted a graph neural network (GNN) model to learn streamflow spatiotemporal dynamics at the basin scale. In particular, we sought to (a) illuminate the role of physics-based connectivity in GNN river network learning, and (b) adapt and investigate the efficacy of a graph-based data fusion technique. The multistage GNN modeling framework was demonstrated using the National Water Model (NWM) reanalysis data for a snow-dominated watershed located in the headwaters region of the Upper Colorado River Basin, U.S. Two sets of experiments were conducted, one using the river network in NWM and the other using a much coarser network correspond to subbasin outlets. This talk will give an in-depth look of the experimental results, as well as providing further insights on basin-scale, streamflow dynamics modeling.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H45L1533S