Martel et al. (2025) Exploring the ability of LSTM-based hydrological models to simulate streamflow time series for flood frequency analysis

Identification

• Journal: Hydrology and earth system sciences
• Year: 2025
• Date: 2025-10-06
• Authors: Jean‐Luc Martel, Richard Arsenault, Richard Turcotte, Mariana Castañeda-González, François Brissette, William F. Armstrong, Edouard Mailhot, Jasmine Pelletier-Dumont, Simon Lachance‐Cloutier, Gabriel Rondeau‐Genesse, Louis‐Philippe Caron
• DOI: 10.5194/hess-29-4951-2025

Research Groups

• Hydrology, Climate and Climate Change (HC3) laboratory, École de technologie supérieure, Montréal, Canada
• Direction principale de l’expertise hydrique (DPEH), Ministère de l’Environnement et de la Lutte contre les changements climatiques, de la Faune et des Parcs (MELCCFP), Québec, Canada
• Ouranos, Montréal, Canada

Short Summary

This study investigates six methods to improve the peak streamflow simulation skill of Long Short-Term Memory (LSTM) models for flood frequency analysis (FFA) in ungauged catchments. It demonstrates that hybrid LSTM-hydrological model implementations can simulate peak streamflow as well as, or better than, traditional distributed hydrological models.

Objective

• To determine whether LSTM-based hydrological models can generate accurate peak streamflow predictions essential for effective flood risk management and reliable flood frequency analysis (FFA) in ungauged catchments.

Study Configuration

• Spatial Scale: 88 catchments in southern Quebec, Canada, with drainage areas ranging from 5.0 x 10^7 m² to 5.0 x 10^10 m².
• Temporal Scale: Daily time step, with data records covering periods between 1979 and 2023, ensuring at least 10 years of observed streamflow data per catchment.

Methodology and Data

• Models used:
  • Long Short-Term Memory (LSTM) networks (7 variants: LSTM-Base, LSTM-Meteo, LSTM-HYDROTEL, LSTM-Multi-head, LSTM-Oversampling, LSTM-Donors, LSTM-Combined).
  • HYDROTEL (semi-distributed physically-based hydrological model).
  • Generalized Extreme Value (GEV) and Gumbel distributions for Flood Frequency Analysis (FFA).
• Data sources:
  • Meteorological data:
    • Quebec government in-house gridded precipitation and air temperature dataset (0.1° resolution, daily, 1979–2017).
    • ERA5 reanalysis product (0.25° resolution, hourly aggregated to daily, 1979–2023) including maximum/minimum air temperature [°C], total precipitation [m], rainfall [m], snowfall [m], snowmelt [m], snow water equivalent [m], dew point temperature [°C], wind velocity on east–west axis [m s⁻¹], wind velocity on north–south axis [m s⁻¹], wind speed [m s⁻¹], evaporation [m], downward surface solar radiation [J m⁻²], and surface pressure [Pa].
  • Hydrometric data: Daily average streamflow [m³ s⁻¹] from 88 stations in Quebec (1979–2017).
  • Catchment descriptors: 24 descriptors (8 shape/geographic, 7 land use, 9 hydrometeorological) extracted using the PAVICS-Hydro platform.
  • Additional donor catchments: Data from 500 catchments from the HYSETS database for the LSTM-Donors model.

Main Results

• LSTM models generally achieved better Kling-Gupta Efficiency (KGE) for overall streamflow simulation compared to the HYDROTEL model.
• LSTM-based models demonstrated the ability to simulate peak streamflow (NRMSE Qx1day) as well as or better than the distributed hydrological model.
• The "LSTM-Combined" model, which integrated multi-head attention, additional meteorological data, and HYDROTEL simulations, exhibited the best performance, outperforming HYDROTEL in 95.5% of catchments for KGE and 72.7% for NRMSE Qx1day.
• Incorporating HYDROTEL simulations as an input was the most impactful individual strategy for improving both KGE and NRMSE Qx1day.
• Oversampling peak streamflow events led to worse NRMSE Qx1day results compared to the baseline LSTM model.
• Adding 500 donor catchments improved both KGE and NRMSE Qx1day.
• Catchment drainage area and the length of the observational record did not significantly correlate with LSTM model performance.
• Flood Frequency Analysis (FFA) using the Gumbel distribution showed that both LSTM-Combined and HYDROTEL model simulations produced FFAs within the uncertainty bounds of observations. The LSTM-Combined model tended to match or underestimate the observed distribution, while HYDROTEL showed both overestimation and underestimation.

Contributions

• Systematically investigates and compares six strategies to enhance LSTM-based hydrological models for peak streamflow simulation and flood frequency analysis (FFA) in ungauged catchments.
• Demonstrates that hybrid LSTM-hydrological model approaches, particularly those incorporating simulations from traditional hydrological models as input, significantly improve LSTM performance for peak streamflow, often surpassing standalone process-based models.
• Provides empirical evidence for the value of integrating "expert" knowledge (e.g., hydrological model outputs) into deep learning frameworks for hydrological modeling.
• Offers insights into the limitations of deep learning models for rare event simulation (e.g., peak flows) and explores methods to mitigate these, contributing to the development of more robust flood risk management tools.

Funding

• INFO-Crue research program (project no. 711500)

Citation

@article{Martel2025Exploring,
  author = {Martel, Jean‐Luc and Arsenault, Richard and Turcotte, Richard and Castañeda-González, Mariana and Brissette, François and Armstrong, William F. and Mailhot, Edouard and Pelletier-Dumont, Jasmine and Lachance‐Cloutier, Simon and Rondeau‐Genesse, Gabriel and Caron, Louis‐Philippe},
  title = {Exploring the ability of LSTM-based hydrological models to simulate streamflow time series for flood frequency analysis},
  journal = {Hydrology and earth system sciences},
  year = {2025},
  doi = {10.5194/hess-29-4951-2025},
  url = {https://doi.org/10.5194/hess-29-4951-2025}
}