Journal Article

Development of a Distributed Physics‐Informed Deep Learning Hydrological Model for Data‐Scarce Regions

Lei, Huimin, Yang, Jingjing, Zhong, Liangjin

2024

Overview

Climate change has exacerbated water stress and water‐related disasters, necessitating more precise streamflow simulations. However, in the majority of global regions, a deficiency of streamflow data constitutes a significant constraint on modeling endeavors. Traditional distributed hydrological models and regionalization approaches have shown suboptimal performance. While current deep learning (DL)‐related models trained on large data sets excel in spatial generalization, the direct applicability of these models in certain regions with unique hydrological processes can be challenging due to the limited representativeness within the training data set. Furthermore, transfer learning DL models pre‐trained on large data sets still necessitate local data for retraining, thereby constraining their applicability. To address these challenges, we present a physics‐informed DL model based on a distributed framework. It involves spatial discretization and the establishment of differentiable hydrological models for discrete sub‐basins, coupled with a differentiable Muskingum method for channel routing. By introducing upstream‐downstream relationships, model errors in sub‐basins propagate through the river network to the watershed outlet, enabling the optimization using limited downstream streamflow data, thereby achieving spatial simulation of ungauged internal sub‐basins. The model, when trained solely on the downstream‐most station, outperforms the distributed hydrological model in streamflow simulation at both the training station and upstream held‐out stations. Additionally, in comparison to transfer learning models, our model requires fewer gauge stations for training, but achieves higher precision in simulating streamflow on spatially held‐out stations, indicating better spatial generalization ability. Consequently, this model offers a novel approach to hydrological simulation in data‐scarce regions, especially those with poor hydrological representativeness. Plain Language Summary Climate change leads to more water shortages and disasters, requiring better streamflow predictions. Yet, a big hurdle in dealing with this issue is the lack of streamflow data across many parts of the world. Traditional physics‐based distributed hydrological models and current deep learning (DL) models have their limitations, especially for regions with unique hydrological processes and limited observations. To address these challenges, we developed a new tool combining physics‐informed DL and a traditional river routing model based on the distributed framework. The model divides the region into sub‐basins, where a physics‐informed DL rainfall‐runoff model calculates runoff generation, and a physics‐informed DL routing model computes the movement of water within each subunit toward the river. Model errors propagate downstream through the river network, thus requiring only a small amount of downstream data to optimize all sub‐basin models and effectively simulate internal unmonitored sub‐basins. When solely using the downstream‐most discharge stations for training, our model outperforms the traditional physics‐based distributed hydrological model. In addition, our approach requires less training data than transfer learning, while achieving higher spatial generalization accuracy. In summary, our model provides a new way to simulate streamflow in data‐scarce regions with unique processes. Key Points A distributed physics‐informed deep learning hydrological model was proposed for data‐scarce regions The new model outperforms the traditional distributed hydrologic model in simulating streamflow in upstream held‐out stations Our model requires less data for training but performs better than the transfer learning model in spatial generalization

• Publisher: John Wiley & Sons, Inc,Wiley
• Subject: Basins / Climate change / Climate models / Climate prediction / Constraint modelling / data collection / Datasets / data‐scarce regions / Deep learning / Disasters / distributed physics‐informed deep learning / Downstream / Errors / Hydrologic data / Hydrologic models / Hydrologic processes / Hydrology / Machine learning / Observational learning / Outlets / Physics / physics‐informed deep learning / Rainfall-runoff relationships / Regions / River networks / Rivers / Runoff / Simulation / Spatial data / Spatial discrimination learning / Stream discharge / Stream flow / Streamflow data / streamflow simulation / Training / Transfer learning / ungauged basins / Upstream / water / Water shortages / Water stress / watersheds