Comparing single and multi-objective hydrologic model calibration considering reservoir inflow and streamflow observations

Abstract

Calibration techniques were investigated on how to best optimize a 149 parameter, distributed hydrological model of the Lake of the Woods – Rainy Lake (LOWRL) watershed. Single objective calibrations based only on streamflow, only on reservoir inflows or average performance of both observation types, optimized using the Dynamically Dimensioned Search (DDS) algorithm, were compared with a multi-objective optimization approach with both observation types using the Pareto Archived DDS (PADDS) algorithm. Results from synthetic calibration tests against a known solution showed that PADDS was able to repeatedly find solutions with streamflow and reservoir inflow Nash-Sutcliffe coefficients of more than 0.95 and 0.99 using 2000 and 8000 model evaluations, respectively, demonstrating the effectiveness of PADDS on a limited calibration budget. When the LOWRL model was calibrated to actual observations with PADDS using 2000 evaluations, the algorithm repeatedly returned solutions with validation period streamflow and reservoir inflow Nash-Sutcliffe coefficients of approximately 0.71 and 0.87, respectively. Results demonstrate the capabilities of PADDS to reasonably calibrate a large dimensional hydrologic model on a restricted budget of 2000 model evaluations and highlight the importance of calibrating to both reservoir inflows and streamflows simultaneously. Considering the comparative results under multiple calibration trials, the multi-objective formulation solved by PADDS is shown to generate equivalent quality results as a weighted single objective approach solved by DDS (averaging reservoir inflow and streamflow calibration objectives).