Improving Distributed Hydrological Model Robustness Through Multi‐Variable Calibration

ABSTRACT Hydrological models are essential tools for estimating water balance components, evaluating management scenarios, and assessing extreme events. However, models struggle to adequately simulate conditions different to those used for calibration, such as changes in land uses or future climate scenarios. This study aims to enhance the robustness of distributed hydrological models through multi‐variable calibration and dynamically representing land use change. A novel Bayesian Regression Robustness Assessment Test (BRRAT) is introduced to quantify improvements to robustness, defined as independence between inputs and model errors. The Australian Water Resources Assessment Landscape (AWRA‐L) model was applied to catchments in southern Australia. Calibration datasets included streamflow only, as well as including groundwater levels, soil moisture and leaf area index (LAI). Limited trade‐off between the accuracy of the modelled streamflow was found when using a multi‐variable calibration approach, particularly for the higher runoff catchments considered. The BRRAT found improved model robustness through multi‐variable calibration due to improved representation of the internal catchment processes by including the additional dataset in the calibration. Scenarios of land use and climate change demonstrated model extrapolation. Results from the streamflow only calibration were inconsistent with the expected changes in the water balance, which was not the case for the multi‐variable calibration. Changes in water balance components of recharge and runoff were greater for the model calibrated to multiple datasets due to increased modelled evapotranspiration. The study demonstrates the importance of considering multiple variables in hydrological model calibration to achieve more accurate and robust predictions. The BRRAT provides an approach to systematically test for improvements in modelled robustness with changes in calibration approach from one model simulation, without requiring time‐consuming multiple split‐sample recalibration. This calibration and evaluation approach improves the confidence in hydrological models' ability to inform water management decisions, particularly in regions experiencing significant environmental changes.