Performance of nearest neighbour metrics for pluvial flood nowcasts in urban catchments

Abstract

Urban flood warning systems need fast response times between the rainfall forecast and the flood alarm. Flood forecasts from physically based numerical models usually need much computation time. Flood forecasts based on databases from previous events or pre-simulated events can speed up the process of decision making. This work introduces and compares four distance metrics for temporal rainfall patterns used in a nearest neighbour based forecast system for dynamic water levels and velocities during pluvial floods. The system uses a database of 960 pre-calculated flood events. The performance of each metric is evaluated by analysis of time-series of flow variables. For the error quantification,a procedure to find a small number of representative locations in an urban catchment is described. A new approach to quantify the similarity of dynamic flow fields is introduced, which makes use of particle transport tracking. The four distance metrics are tested on forecast for four exemplary pluvial flood events resulting from rainfall of durations between 15 and 50 min and return-periods between 10 and 100 years. The best metric is based on the temporal precipitation pattern and takes the response time of the drainage system into account. If the pattern has a very short duration, a simpler characterisation-metric can be used, which takes the total volume, peak intensity and peak position into account.