Hybrid models generalize better to warmer climate conditions than process-based and purely data-driven models

Jan Philipp Bohl, Raul R. Wood, Corinna Frank, Paul C. Astagneau, Jonas Peters, Manuela I. Brunner2025 · DOI

TL;DR

While deep learning models (LSTMs) are currently the "gold standard" for predicting streamflow in unknown regions, they struggle to adapt to the shifting temperatures of a warming planet. This study demonstrates that hybrid models—which use neural networks to dynamically tune the parameters of physical "bucket" models—provide the best of both worlds: the high accuracy of AI with the robust climate-generalizability of traditional physics.

What Problem Does This Paper Address?

Hydrologists are increasingly turning to Long Short-Term Memory (LSTM) networks to predict rainfall-runoff, especially in "ungauged basins" where historical data is missing. While LSTMs excel at spatial generalization (predicting what happens in a new place), their ability to handle temporal generalization (predicting what happens in a future, warmer climate) is a major question mark.

If a model is trained on data from a relatively cool period (e.g., 1960–1990), can we trust it to perform accurately during a 21st-century heatwave? Purely data-driven models are notorious for failing when they encounter "out-of-distribution" data. This paper investigates whether adding a layer of physical conceptualization—creating a hybrid model—can prevent these models from "breaking" as the climate warms.

Methods

The researchers compared three distinct modeling approaches across 918 catchments in Central Europe:

  1. Purely Data-Driven (LSTM): A standard recurrent neural network that learns the relationship between weather and streamflow from scratch.
  2. Conceptual (HBV): The "Hydrologiska Byråns Vattenbalansavdelning" model, a traditional process-based model that represents the environment as a series of interconnected "buckets" (snow, soil, groundwater).
  3. Hybrid (LSTM-HBV): A sophisticated middle ground where an LSTM is used to dynamically predict the parameters (like the runoff coefficient or evapotranspiration factor) of the HBV model for every single day.

To test climate robustness, they used a Differential Split Sample Test (DSST). They trained the models on a "cold" historical period (1960–1990) and then evaluated them on a significantly "warm" period (2000–2023) and an even more extreme "warmest 5 years" subset.

Datasets

  • Catchments: 918 near-natural catchments across the European Alps and surrounding lowlands (Germany, Austria, Switzerland, and France).
  • Meteorology: Daily precipitation, temperature, and radiation data from the E-OBS v29.0 gridded dataset.
  • Static Attributes: Land use, geology, and topography data derived from the Hydroatlas and CARAVAN datasets.
  • Snow Benchmarks: Gridded Snow-Water Equivalent (SWE) data from SNOWGRID-CL (Austria) and OSHD-CLQM (Switzerland) were used to validate internal model states.

Key Insights and Ideas

  • The Accuracy-Generalizability Trade-off: The "best" model isn't always the most robust. While the LSTM was the most accurate during the period it was trained on, it saw the steepest decline in performance when temperatures rose.
  • Physics as a Regularizer: The hybrid model acts as a "constrained" version of the LSTM. By forcing the neural network to work within the mass-balance framework of the HBV model, the researchers prevented the AI from making physically impossible predictions when faced with unseen heat.
  • The "Snow Problem": Surprisingly, in snow-dominated catchments, all models (including the conceptual ones) suffered. This suggests that our traditional "degree-day" physical formulas for snowmelt might be too simple to capture the complex dynamics of a rapidly warming Alpine environment.
  • Data Bias Correction: LSTMs showed a unique ability to "learn" and correct for systematic errors in input data (like precipitation undercatch), which traditional physical models cannot do without manual calibration.

Key Results

  • Overall Performance: In the cold training period, the LSTM led with a median Kling-Gupta Efficiency (KGE) of 0.76, followed by the Hybrid (0.72) and regional HBV (0.60).
  • Climate Shift: When moving to the warm period, the LSTM’s accuracy dropped by 0.05 KGE points. The Hybrid and HBV models were much more stable, dropping by only 0.02.
  • The Hybrid Sweet Spot: In the "warmest" years, the Hybrid model's accuracy was statistically indistinguishable from the LSTM, but it provided more reliable estimates for extreme events like droughts.
  • Extreme Events: Hybrid models were particularly effective at simulating drought volumes, outperforming both purely data-driven and purely conceptual models in warmer conditions.

Limitations and Future Work

The authors acknowledge that the HBV model's simplicity is a double-edged sword; while it provides a robust framework, its basic snow and evapotranspiration modules might limit the hybrid model's ultimate potential.

Future research should explore:

  • More complex physical cores: Swapping HBV for more detailed energy-balance models.
  • Geographic expansion: Training on catchments from naturally warmer regions to "teach" the models about future conditions.
  • Loss Function Innovation: Modifying how the AI is trained to specifically prioritize "time-invariant" features, further improving its ability to handle climate shifts.

Summary updated

Abstract

Deep-learning based rainfall-runoff models, in particular long short-term memory networks (LSTM), have been shown to outperform traditional hydrological models at various tasks, both when used as purely data-driven models and when combined with process-based models in a hybrid setting. These tasks include predictions in ungauged basins (PUB) and regions (PUR), tasks which have traditionally been challenging for conceptual hydrological models. While the spatial generalizability of deep-learning based models has received a lot of attention, it is less clear how they generalize to unseen and warmer climate conditions, i.e. how suitable these models are for hydrological climate impact studies. To address this research gap, we assess the ability of three types of models including (1) fully data-driven (LSTMs), (2) conceptual (Hydrologiska Byråns Vattenbalansavdelning (HBV)), and (3) hybrid (LSTM-HBV) models to simulate streamflow under conditions warmer than those used to train the models by running a differential split sample test. That is, we trained the models using data from the historical period 1960–1990 and evaluated them on both data of this period as well as of the warmer period 2000–2023. We find that LSTMs, while being the most accurate during the 1960–1990 period, have inferior generalizability to the warm period compared to the hybrid and conceptual models. In addition, we show that when generalizing to the warm period, hybrid models have similar accuracy as LSTMs, independently of whether the entire streamflow distribution or extreme events such as floods and droughts are considered. However, for snow-dominated catchments, all models suffer from similar reductions in accuracy when simulating streamflow under unseen climate conditions and the LSTM is the most accurate model for all periods. A detailed look at the snowmelt simulations of the hybrid and conceptual model suggests that better process-representation might be needed to accurately capture the dynamics of snow-melt and -accumulation processes, which are highly sensitive to changes in temperature. We conclude that the hybrid models effectively combine the high accuracy of LSTMs when predicting in ungauged basins with the good generalizability under changes in climate of conceptual hydrological models. This makes them a suitable choice for hydrological climate change impact assessments, particularly in ungauged basins.
BibTeX
@misc{bohl2025hybrid,
  author = {Jan Philipp Bohl and Raul R. Wood and Corinna Frank and Paul C. Astagneau and Jonas Peters and Manuela I. Brunner},
  title = {{Hybrid models generalize better to warmer climate conditions than process-based and purely data-driven models}},
  year = {2025},
  doi = {10.5194/egusphere-2025-5201},
}
Figure