Li et al. (2025) Ensembling differentiable process-based and data-driven models with diverse meteorological forcing datasets to advance streamflow simulation

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-12-01
• Authors: Peijun Li, Yalan Song, Ming Pan, Kathryn Lawson, Chaopeng Shen
• DOI: 10.5194/hess-29-6829-2025

Research Groups

• Civil and Environmental Engineering, The Pennsylvania State University, University Park, PA, USA
• Center for Western Weather and Water Extremes, Scripps Institution of Oceanography, University of California San Diego, La Jolla, CA, USA

Short Summary

This study systematically evaluates and utilizes ensembles of a data-driven Long Short-Term Memory (LSTM) network and a physics-informed differentiable HBV ($\delta$HBV) model with diverse meteorological forcing datasets to advance streamflow simulation. The research demonstrates that cross-model-type ensembles consistently outperform single-model approaches and set new accuracy benchmarks, particularly enhancing spatial generalization due to complementary error characteristics and the structural constraints of $\delta$HBV.

Objective

• Will a cross-model-type ensemble of LSTM and $\delta$HBV improve deterministic streamflow prediction more than a within-class ensemble?
• Is it better to use multiple forcings in one model or to ensemble multiple models, each with a different forcing input?
• Do process-based equations bring unique value to an ensemble, especially in terms of spatial generalizability?

Study Configuration

• Spatial Scale: 531 river basins across the conterminous United States, derived from the CAMELS dataset. Basin sizes range from 1 to 25 800 square kilometers (median: 335 square kilometers).
• Temporal Scale: Daily temporal resolution for meteorological forcing data and streamflow observations. Evaluations conducted across different periods (temporal test) and for ungauged basins/regions (PUB/PUR tests) over multi-year periods.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) network (data-driven)
  • Differentiable HBV ($\delta$HBV) model (physics-informed machine learning, specifically $\delta$HBV1.1p)
  • Ensemble averaging of individual model outputs
• Data sources:
  • CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) dataset: Streamflow observations, static basin attributes (e.g., basin area, topography, climate, soil texture, land cover, geology).
  • Meteorological forcing datasets: Daymet, North American Land Data Assimilation System (NLDAS), and Maurer (daily precipitation, temperature, vapor pressure, surface radiation).
  • Potential evapotranspiration calculated using the Hargreaves method.

Main Results

• Cross-model-type ensembles (LSTM + $\delta$HBV) consistently surpassed single-model approaches and within-class ensembles across all temporal and spatial generalization tests.
• New benchmark records were established on the CAMELS dataset, achieving median Nash-Sutcliffe model efficiency coefficients (NSE) of approximately 0.83 for the temporal test, 0.79 for the ungauged basin test (PUB), and 0.70 for the ungauged region test (PUR).
• Ensembling models trained on individual meteorological forcing datasets (e.g., LSTM$^{123}$) yielded higher performance (NSE of 0.8082) than feeding multiple forcing datasets simultaneously into a single LSTM model (LSTM$_{multi}$, NSE of 0.7974).
• The $\delta$HBV model significantly improved ensemble performance for spatial interpolation (PUB) and, more prominently, for spatial extrapolation (PUR), demonstrating the value of its structural constraints.
• LSTM and $\delta$HBV exhibited distinct error characteristics that complemented each other, leading to improved high-flow and low-flow metrics in ensembles.
• The most substantial performance improvements from ensembling were observed in the Great Plains and midwestern US, regions historically challenging for hydrological models.

Contributions

• Provides a systematic evaluation of ensembling highly structurally different hydrological models (data-driven and physics-informed) under comprehensive spatiotemporal generalization tests.
• Establishes new state-of-the-art performance benchmarks for streamflow simulation on the CAMELS dataset using cross-model-type ensembles.
• Challenges the conventional approach of fusing multiple forcing datasets into a single data-driven model, demonstrating that ensembling models trained on separate forcings is more effective.
• Highlights the critical role of physics-informed models like $\delta$HBV in providing valuable structural constraints that significantly enhance spatial generalization capabilities of ensembles, particularly in ungauged regions.
• Advances the understanding of how to effectively leverage diverse model types and multi-source datasets to improve streamflow simulations across various hydrological scenarios.

Funding

• Office of Biological and Environmental Research of the U.S. Department of Energy (contract no. DESC0016605)
• California Department of Water Resources Atmospheric River Program Phase III (Grant 4600014294)
• Cooperative Institute for Research to Operations in Hydrology (CIROH) through the National Oceanic and Atmospheric Administration (NOAA) Cooperative Agreement (Grant no. NA22NWS4320003)

Citation

@article{Li2025Ensembling,
  author = {Li, Peijun and Song, Yalan and Pan, Ming and Lawson, Kathryn and Shen, Chaopeng},
  title = {Ensembling differentiable process-based and data-driven models with diverse meteorological forcing datasets to advance streamflow simulation},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-6829-2025},
  url = {https://doi.org/10.5194/hess-29-6829-2025}
}