Assessing urban hydrologic prediction accuracy through event reconstruction

Quantifying hydrologic prediction error is of fundamental importance to feasible operation of a real-time flood forecasting system customized for specific locations. Reconstruction of a series of storm events using gauge-adjusted radar provides information on prediction accuracies that may be achievable in real-time operations. The subject of this paper is the examination of hydrologic prediction uncertainty for a system providing site-specific flood forecasts in a heavily urbanized area. Reconstruction of Tropical Storm Allison that occurred in June 2001 and other events are used to calibrate and validate a fully distributed hydrologic model that is used for real-time continuous forecasting operations. Hydrologic prediction accuracy expected during real-time forecasting of urban flooding is shown to be sensitive to radar rainfall bias. Controlling for bias in the radar input for five storms, achievable model accuracy is approximately 11.8% in peak discharge, 12 min in timing, and 11.1% in runoff volume for a 260 km2 urban catchment. High-accuracy rising limb agreement is achieved through differential adjustment of overland and channel hydraulic roughness. Verification during a recent real-time event illustrates the need for bias adjustment of radar in real-time to achieve acceptable accuracies. Once the radar is gauge-corrected, remaining random error does not correlate strongly with the error in streamflow predictions. This lack of correlation indicates that the random error measured by radar-gauge comparison is diminished, not amplified, when simulated runoff is compared with observed streamflow. Performance evaluation during operations will be accomplished in the future as sufficient events become available for reliable statistical evaluation.