Assessing spatial connectivity effects on daily streamflow forecasting using Bayesian-based graph neural network.

Area of Science:

• Hydrology
• Environmental Science
• Data Science

Background:

• Data-driven models show promise in streamflow prediction but often overlook spatial connectivity.
• Existing models focus on input features and structure, neglecting the impact of hydrological station relationships.

Purpose of the Study:

• To develop a novel graph neural network model that accounts for spatial connectivity in streamflow forecasting.
• To assess the impact of basin network structure on daily streamflow prediction accuracy and reliability.

Main Methods:

• Constructed a basin network using graph-structured data to represent spatial connectivity between hydrological stations.
• Proposed a variational Bayesian edge-conditioned graph convolution model integrating edge-conditioned convolutions and Bayesian inference.
• Applied the model to forecast next-day streamflow in the Yangtze River Basin, validated against six comparative models and three experimental groups.

Main Results:

• The proposed graph neural network model achieved high deterministic prediction accuracy (NSE ≈ 0.980, RMSE ≈ 1362.7, MAE ≈ 745.8).
• The model demonstrated excellent probabilistic prediction reliability (ICPC ≈ 0.984, CRPS ≈ 574.1).
• Results confirmed that incorporating spatial connectivity in basin networks enhances forecasting performance.

Conclusions:

• Establishing appropriate connectivity and identifying connection relationships in basin networks is crucial for improving streamflow forecasting.
• The variational Bayesian edge-conditioned graph convolution model effectively captures spatial effects for superior deterministic and probabilistic streamflow predictions.
• This research offers a new approach to hydrological modeling by emphasizing the significance of spatial relationships within a basin.