CoChat

Ensembling differentiable process-based and data-driven models with diverse meteorological forcing datasets to advance streamflow simulation

Open access
Li, Peijun, Song, Yalan, Pan, Ming, Lawson, Kathryn, Shen, Chaopeng · 2025 · Hydrology and Earth System Sciences

Published version of a preprint

The DOI matches its Crossref record; 85 of the 85 references it cites were found in a registry or an index.

Abstract

Streamflow simulations produced by different hydrological models exhibit distinct characteristics and can provide valuable information when ensembled. However, few studies have focused on ensembling simulations from models with significant structural differences and evaluating them under both temporal and spatial tests. Here we systematically evaluated and utilized the simulations from two highly different models with great performances: a purely data-driven long short-term memory (LSTM) network and a physics-informed machine learning (“differentiable”) HBV (Hydrologiska Byråns Vattenbalansavdelning) model (δHBV). To effectively display the features of the two models, multiple forcing datasets are employed. The results show that the simulations of LSTM and δHBV have distinct features and complement each other well, leading to better Nash-Sutcliffe model efficiency coefficients (NSE) and improved high-flow and low-flow metrics across all spatiotemporal tests, compared to within-class ensembles. Ensembling models trained on a single forcing outperformed a single model using fused forcings, challenging the paradigm of feeding all available data into a single data-driven model. Most notably, δHBV significantly enhanced spatial interpolation when incorporated into LSTM, and provided even more prominent benefits for spatial extrapolation where the LSTM-only ensembles degraded significantly, attesting to the value of the structural constraints in δHBV. These advances set new benchmark records on the well-known CAMELS (Catchment Attributes and Meteorology for Large-sample Studies) hydrological dataset, reaching median NSE values of ∼ 0.83 for the temporal test (densely trained scenario), ∼ 0.79 for the ungauged basin test (PUB, Prediction in Ungauged Basins), and ∼ 0.70 for the ungauged region test (PUR, Prediction in Ungauged Regions). This study advances our understanding of how various model types, each with distinct mechanisms, can be effectively leveraged alongside multi-source datasets across diverse scenarios.

Overview

CoChat Summary

TL;DR

This study ensembles a purely data-driven LSTM with a differentiable, physics-informed HBV model (δHBV), each driven by three different meteorological forcing datasets, to test whether cross-model-type ensembles beat within-class ensembles for streamflow simulation. The cross-type ensemble set new CAMELS benchmarks, with δHBV contributing the most value under spatial generalization.

Key findings

  • Cross-model-type ensembles consistently outperformed within-class ensembles across all tests. With a single forcing, median NSE rose from ∼0.735 (LSTM) to ∼0.79 when δHBV was added; adding δHBV to the forcing-ensemble LSTM123 (∼0.808) raised it to ∼0.818.
  • The most diverse ensemble reached median NSE of ∼0.83 (temporal), ∼0.79 (PUB), and ∼0.70 (PUR), surpassing the previous ∼0.82 temporal benchmark.
  • Ensembling models each trained on one forcing beat feeding all forcings into one model: LSTM123 = 0.8082 vs. LSTMmulti = 0.7974.
  • δHBV's benefit grew with generalization difficulty: adding it to LSTM ensembles improved median NSE by 0.02–0.03 for PUB and 0.04–0.07 for PUR (e.g., LSTM123 ∼0.656 vs. (LSTM+δHBV)123 ∼0.701 in PUR).
  • δHBV had lower spread (median 0.61) than LSTM (mean 0.72), but combining them increased ensemble spread and diversity.

Methods

  • Two model types (LSTM, following Kratzert et al. 2021 hyperparameters; δHBV1.1p, following Feng et al. 2023b) were each trained with three CAMELS forcing datasets (Daymet, NLDAS, Maurer), yielding six single-forcing simulations; an LSTMmulti trained on all three forcings simultaneously was added for comparison.
  • Evaluation used the 531-basin CAMELS subset across three scenarios: temporal test, PUB (10 spatially stratified holdout rounds), and PUR (7 disjoint regional holdout rounds); all simulations were repeated with three random seeds (10 seeds in a supplementary check).
  • Metrics were NSE, KGE, PBIAS, RMSE, plus high-flow (top 2 %), low-flow (bottom 30 %), and mid-range RMSE; ensembles were formed by simple averaging after testing weighting strategies, with spread used to assess diversity.

Limitations

  • Stated: no δHBVmulti model was trained (deferred to Sawadekar et al. 2025), and "seed" combinations of δHBV were not examined because its low variance was expected to yield little benefit; averaging was used rather than dynamic/adaptive weighting, though optimized weights suggested further gains are possible.
  • Stated: both models perform poorly in regions with strong anthropogenic influence (dams, arid climates, heterogeneous geology) in the midwestern and western CONUS; δHBV is more sensitive to precipitation bias than LSTM and fits different forcing datasets less well; ensemble computational cost may be a constraint for large-scale or real-time use.
  • Evident from the text: the claim that models extrapolate well under warming is not directly tested, since the authors note no substantially warmed dataset is available; robustness checks (10 % precipitation perturbation, sample size, parameter choices, lookback windows) were limited to the temporal test.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.5194/hess-29-6829-2025

Cites

85 of 85 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

0 Valid · 2 Not valid · 0 Not checked · 10 Own results and interpretations

7 more claims

Create a free account to see all 12 claims and every source verdict.

See all claims — free account

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Citations

Cites

Show all 85 references

Research works printed but not linked: 2

  1. Aghakouchak, A. and Habib, E.: Application of a Conceptual Hydrologic Model in Teaching Hydrologic Processes, International Journal of Engineering Education, 26, 963–973, 2010.
  2. Bergström, S.: Development and application of a conceptual runoff model for Scandinavian catchments, PhD Thesis, Swedish Meteorological and Hydrological Institute (SMHI), Norköping, Sweden, urn:nbn:se:smhi:diva-5738, 1976.

Metadata

Authors
Li, P., Song, Y., Pan, M., Lawson, K., & Shen, C.
Affiliations
  • Pennsylvania State University (US)
  • Scripps Institution of Oceanography (US)
Corresponding author
  • Li, Peijun
  • Shen, Chaopeng
Venue
Hydrology and Earth System Sciences
Published
2025-12-01
Volume
29
Issue
23
Pages
6829-6861
ISSN
  • 1027-5606 (ISSN-L)
  • 1607-7938
OpenAlex
W4416842778
Cited by
7
References
80
Citations by year
2026: 7
FWCI
1.93
Type
journal article
Language
en
Open access
Locations
  • Hydrology and earth system sciences (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
  • National Oceanic and Atmospheric Administration (A23-0252-S002)
  • Biological and Environmental Research (DESC0016605)
  • Department of Water Resources (4600014294)
Topics
  • Physical Sciences › Environmental Science › Environmental Engineering › Hydrological Forecasting Using AI
  • Physical Sciences › Environmental Science › Water Science and Technology › Hydrology and Watershed Management Studies
  • Physical Sciences › Environmental Science › Management, Monitoring, Policy and Law › Environmental Monitoring and Data Management
Keywords
streamflow simulation, model ensembling, physics-informed machine learning, LSTM, spatial extrapolation, ungauged basins, ungauged areas, camels

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

A language model drew the claims from the paper on 2026-09-26 (prompt extract-v2).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Ensembling differentiable process-based and data-driven models with diverse meteorological forcing datasets to advance streamflow simulation. https://app.cochat.ai/papers/10.5194/hess-29-6829-2025
The paper
Li, P., Song, Y., Pan, M., Lawson, K., & Shen, C. (2025). Ensembling differentiable process-based and data-driven models with diverse meteorological forcing datasets to advance streamflow simulation. Hydrology and Earth System Sciences. https://doi.org/10.5194/hess-29-6829-2025

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.