A Proof of Concept for Improving Estimates of Ungauged Basin Streamflow
Via an LSTM-Based Synthetic Network Simulation Approach
Abstract
This study introduces a machine learning approach to address the
critical challenge of limited real-time flow data in river basins,
particularly for calibrating large-scale hydrologic models. These models
often rely on uncertain parameter transfers from gauged to ungauged
regions, hindering accurate predictions. To mitigate this, we propose
utilizing long short-term memory (LSTM) networks to estimate historic
streamflow in ungauged watersheds, effectively creating “surrogate
gauges.” Our innovative method treats watersheds as interconnected
systems, leveraging downstream flow information to constrain and improve
upstream flow estimates. By training and testing the LSTM on synthetic
networks of conceptual ”leaky bucket” watershed models, we demonstrate
its ability to outperform traditional methods. Notably, by representing
watersheds as paired upstream-downstream networks, this concept can be
generalized to any ungauged portion of a real basin, enhancing flow
estimates in data-scarce scenarios. This research represents a proof of
concept, highlighting the potential for significantly enhancing
hydrologic modeling in data-scarce regions and providing a scalable
method for watershed network analysis. Future work will focus on
applying the model to real-world basins to validate its performance and
scalability for continental-scale applications. Ultimately, this
research aims to contribute to the development of more accurate and
reliable hydrologic predictions.