Kapoor et al. (2025) QDeepGR4J: Quantile-based ensemble of deep learning and GR4J hybrid rainfall-runoff models for extreme flow prediction with uncertainty quantification

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-10-23
• Authors: Arpit Kapoor, Rohitash Chandra
• DOI: 10.1016/j.jhydrol.2025.134434

Research Groups

• School of Mathematics and Statistics, University of New South Wales, Sydney, NSW, Australia
• Data Analytics for Resources and Environments, Australian Research Council—Industrial Transformation Training Centre, Sydney, NSW, Australia
• Transitional Artificial Intelligence Research Group, School of Mathematics and Statistics, University of New South Wales, Sydney, NSW, Australia

Short Summary

This paper introduces QDeepGR4J, a quantile regression-based ensemble extension of the DeepGR4J hybrid rainfall-runoff model, to quantify uncertainty in multi-step streamflow predictions and identify extreme flow events. The framework significantly improves predictive accuracy and uncertainty interval quality, demonstrating its suitability as an early warning system for floods.

Objective

• To extend the DeepGR4J hybrid rainfall-runoff model using a quantile regression-based ensemble learning framework to quantify uncertainty in streamflow prediction.
• To leverage the predicted uncertainty bounds to identify extreme flow events, potentially leading to flooding.
• To extend the model for multi-step streamflow predictions with uncertainty bounds, aiming to develop a reliable early flood warning system.

Study Configuration

• Spatial Scale: 222 unregulated Australian catchments from the CAMELS-AUS dataset. Detailed evaluations were conducted on all 9 stations in South Australia and 5 selected stations from each of seven Australian states (New South Wales, Northern Territory, Queensland, South Australia, Tasmania, Victoria, Western Australia).
• Temporal Scale: Daily streamflow prediction. Hydrometeorological time-series data from 1980 to 2014 (35 years) was used, with 60% for training and 40% for testing. Multi-step ahead predictions were made with a 3-day forecast horizon.

Methodology and Data

• Models used:
  • GR4J (Génie Rural à 4 paramètres Journalier model): A conceptual rainfall-runoff model.
  • DeepGR4J: A hybrid model enhancing GR4J by replacing its routing component with a deep learning model.
  • QDeepGR4J: The proposed quantile regression-based ensemble of DeepGR4J models.
  • Deep Learning architectures: Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNN), and Multi-Layer Perceptron (MLP).
  • Quantile Regression: Implemented using a tilted loss function to predict specific streamflow quantiles (0.05, 0.50, 0.95).
  • Generalised Extreme Value (GEV) distribution: Used to estimate flood thresholds based on observed annual maximum streamflow.
  • Optimization: Differential Evolution for GR4J parameter calibration and Adam optimizer (learning rate 0.001, 𝛽1 = 0.89, 𝛽2 = 0.97) for neural network training.
• Data sources:
  • CAMELS Australia (CAMELS-AUS) dataset: Provides hydrometeorological time-series data, including streamflow, 12 climate variables (precipitation, evapotranspiration, minimum and maximum temperature, vapor pressure), and 134 catchment attributes (geology, soil, topography).
  • Data processing: Windowing approach with a window size of 7 days, linear interpolation for missing values (if less than 10% missing), and z-score normalization for input and output data.

Main Results

• The LSTM-based QDeepGR4J architecture demonstrated superior performance in median streamflow prediction (RMSE and NSE) and uncertainty interval quality (Interval Score) compared to MLP, RNN, and CNN architectures, particularly in the South Australia region.
• Hybridization significantly improved predictive accuracy (higher NSE, lower RMSE) for both training and test sets across all seven evaluated Australian states for both LSTM and CNN-based QDeepGR4J ensembles compared to their standalone deep learning counterparts.
• QDeepGR4J-LSTM consistently achieved the lowest interval scores on the training set across all states, indicating better capture of data uncertainty with tighter 90% confidence intervals. For test sets, QDeepGR4J-CNN occasionally showed lower interval scores (e.g., Queensland, Western Australia).
• The QDeepGR4J ensemble exhibited a notably high True Positive Rate (TPR) for flood event detection, especially for 3-year and 5-year flood recurrence intervals, outperforming baseline LSTM ensembles. For instance, it achieved up to 1.000 TPR for 3-year and 5-year floods in some stations.
• While effective in capturing extreme events, the model occasionally overestimated streamflow peaks, leading to potential false positive flood alerts. Performance for 7-year and 10-year flood recurrence intervals was lower.
• Multi-step predictions showed a slight increase in uncertainty bounds for longer forecast horizons (e.g., 2 and 3 days ahead).

Contributions

• Introduction of QDeepGR4J, a novel quantile regression-based ensemble framework that extends the DeepGR4J hybrid rainfall-runoff model for comprehensive uncertainty quantification and extreme flow prediction.
• Development of a robust methodology for multi-step ahead streamflow forecasting that provides explicit uncertainty intervals, crucial for hydrological risk assessment.
• Integration of the Generalised Extreme Value (GEV) distribution with predicted uncertainty bounds to derive a qualitative flood risk indicator, enhancing the utility of the model as an early warning system.
• Empirical demonstration of the superior performance of the hybrid QDeepGR4J ensemble over pure deep learning models in terms of both predictive accuracy (RMSE, NSE) and the quality of uncertainty intervals (Interval Score) across diverse Australian catchments.
• Addresses the challenge of poor performance in extreme flow regions, a limitation identified in previous DeepGR4J work, making the model more suitable for flood management.

Funding

• Australian Government through the Australian Research Council Industrial Transformation Training Centre for Data Analytics for Resources and Environments (DARE) (project IC190100031).
• Katana High Performance Computing (HPC) cluster supported by the University of New South Wales (DOI: 10.26190/669X-A286).

Citation

@article{Kapoor2025QDeepGR4J,
  author = {Kapoor, Arpit and Chandra, Rohitash},
  title = {QDeepGR4J: Quantile-based ensemble of deep learning and GR4J hybrid rainfall-runoff models for extreme flow prediction with uncertainty quantification},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134434},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134434}
}