Information-aware selection of calibration events for SWMM models: evidence from the Bellinge benchmark

Calibrating urban drainage models against large wet-weather archives is costly. We developed quality-controlled hydraulic-information portfolio selection (QHIPS) to retain events with complementary multi-sensor SWMM parameter information. In the Bellinge benchmark, 29 qualified events were split chronologically into 20 development and nine holdout events. Morris screening retained five parameters. Surrogate-smoothed gradients guided information-matrix construction and portfolio search; holdout predictions came from direct SWMM simulations. QHIPS achieved the highest development-set regularized log determinant at each tested budget. Selecting five events reduced the archive by 75%; directly comparable downstream calibration-and-verification event-runs and runtime fell by 51.7% and 44.9%, respectively. Median holdout normalized RMSE was 0.8345, versus 0.8508 with all 20 events; the paired difference was −0.0133 (95% bootstrap interval −0.0216 to 0.0114), with the interval including zero. Overall, QHIPS substantially compressed the calibration archive and reduced downstream workload while retaining independently evaluated temporal performance on a similar error scale.