McInerney et al. (2025) Tailored calibration of stochastic weather generators for enhanced hydrological system evaluation

Identification

• Journal: Journal of Hydrology
• Year: 2025
• Date: 2025-12-31
• Authors: David McInerney, Seth Westra, Michael Leonard, Dmitri Kavetski, Mark Thyer, Holger R. Maier
• DOI: 10.1016/j.jhydrol.2025.134894

Research Groups

School of Civil Engineering and Construction, Adelaide University, SA, Australia

Short Summary

Tailored calibration of stochastic weather generators (SWGs) using the Simulated Method of Moments (SMM) significantly improves the capture of critical climate attributes and hydrological responses compared to conventional methods, with the Robust Gauss-Newton (RGN) algorithm proving essential for accurate and efficient optimization in both historical and future climate assessments.

Objective

• Evaluate the benefits of using a tailored calibration approach by comparing its performance with that of a conventional approach, focusing on its ability to improve the performance of system response in historical climates.
• Identify a robust optimization algorithm for the tailored calibration of SWGs that is computationally efficient, provides reliable parameter identification, and reliable estimation of the system response to future climates.

Study Configuration

• Spatial Scale: Three catchments in eastern Australia (Namoi River at North Cuerindi, Abercrombie River at Hadley No. 2 gauge, Cotter River at Gingera), located within 500 km of each other.
• Temporal Scale: Daily rainfall and streamflow data. Historical climate simulations used 50 years and 50 replicates. Perturbed climate simulations used 20 years and 20 replicates.

Methodology and Data

• Models used:
  • Stochastic Weather Generator: WGEN model (daily rainfall).
  • Rainfall-runoff model: GR4J.
  • Calibration methods: Conventional Method of Moments/Maximum Likelihood (MoM/ML) and tailored Simulated Method of Moments (SMM).
  • Optimization algorithms for SMM: Nelder-Mead (NM), Shuffled Complex Evolution (SCE), and Robust Gauss-Newton (RGN).
• Data sources:
  • CAMELS-AUS dataset.
  • Rainfall data: Australian Water Availability Project (AWAP).
  • Potential Evapotranspiration (PET) data: Scientific Information for Land Owners (SILO) project.
  • Streamflow data: Hydrologic Reference Stations project.

Main Results

• Conventional MoM/ML calibration underestimated historic extreme multi-day rainfall by approximately 30 % and extreme streamflow by 30–50 %, while overestimating median rainfall.
• Tailored SMM calibration, with appropriate attribute sets (e.g., multi-day extreme rainfall), nearly eliminated biases in extreme multi-day rainfall and significantly reduced biases in mean and high streamflows (e.g., mean annual flow bias reduced from 28 % to 3–9 %; high flow bias reduced from 50 % to 2–9 %).
• The choice of optimization algorithm for SMM significantly impacted accuracy and efficiency:
  • Nelder-Mead (NM) and Shuffled Complex Evolution (SCE) introduced substantial errors (20 % and 15 % respectively) in simulated climate attributes and resulted in larger uncertainties in streamflow predictions.
  • Robust Gauss-Newton (RGN) achieved significantly higher accuracy (errors around 2 %) in capturing target attributes and reduced uncertainty in future streamflow predictions by up to 75 % compared to other optimizers.
• Computational efficiency: NM was the fastest (2.8 hours for 500 optimization runs) but inaccurate. RGN was moderately fast and accurate (5.2 hours). SCE was the slowest and inaccurate (59 hours).

Contributions

• Provides the first comprehensive comparison of conventional and tailored calibration approaches for SWGs, evaluating both key climate attributes and their influence on system performance using observed climate data.
• Demonstrates that tailored calibration using SMM can overcome limitations of SWGs without modifying their structure, by emphasizing system-critical rainfall attributes.
• Identifies the Robust Gauss-Newton (RGN) algorithm as a superior optimization method for tailored SWG calibration, offering high accuracy in parameter identification and computational efficiency, crucial for both historical and future climate impact assessments.
• Highlights the importance of evaluating simulated climates within system models (e.g., rainfall-runoff models) to ensure that the most relevant aspects of climate are appropriately captured.
• Offers a unified approach for SWG calibration to reliably estimate system responses across historical and future climates.

Funding

Computations were performed on the Phoenix cluster at the University of Adelaide. No specific funding projects or programs were listed.

Citation

@article{McInerney2025Tailored,
  author = {McInerney, David and Westra, Seth and Leonard, Michael and Kavetski, Dmitri and Thyer, Mark and Maier, Holger R.},
  title = {Tailored calibration of stochastic weather generators for enhanced hydrological system evaluation},
  journal = {Journal of Hydrology},
  year = {2025},
  doi = {10.1016/j.jhydrol.2025.134894},
  url = {https://doi.org/10.1016/j.jhydrol.2025.134894}
}