Selecting velocity fields in urban pluvial floods from pre-calculated data

Early warning systems for floods in urban areas should forecast water levels and damage estimation to protect vulnerable regions. To estimate the danger of a flood for buildings and people, the energy of the flood has to be taken into account additionally to the water level. The energy is related to the flow velocity. For directing rescue workers or trace spreading of contaminants through flooded streets, a high resolution of the water’s energy in space and time is required. Direct numerical run-off calculation is too slow for a flood forecast in time. Therefore a database with pre-calculated events is needed and a method to select the water levels and velocity fields that are similar to a forecasted rain event.

We present a method, how to create a real-time forecast based on pre-calculated data. The selection and weighting of the pre-calculated data is based on the precipitation pattern in the flood region. A nearest neighbor approach is applied to find water levels and velocity fields from a database that are similar to the forecasting event. For the ranking of similarity, different new metrics are compared against each other. The quality of the metrics is tested with a new approach of comparing velocity fields on the surface and in the pipe system. Considering both domains is crucial for understanding the complex dynamic flow paths on the surface. An urban catchment of 5 km² with high resolution (~3 m³) triangular surface mesh and connected drainage system is used for a hydrodynamic run-off simulation. The 1D-2D coupled software HYSTEM EXTRAN is used to generate the water levels and velocity fields for strong rainfall events of the past 20 years. More than 900 events with a duration between 15 minutes and 24 hours and return periods between 10 and 100 years were calculated and stored as the “pre-calculated” dataset.

For comparing two events, the mean square error is calculated between the precipitation patterns with different approaches to select the start index and number of intervals. This number depends on the hydraulic response time, the temporal resolution and the length of the reference pattern. The quality of the nearest neighbor selection is quantified using the Nash–Sutcliffe model efficiency coefficient of pipe flow and the root mean square error of water level and velocity in significant surface cells. Additionally, the transport paths of artificial contamination spills are compared between the events to show the reproducibility of velocity fields for each metric.

Results show that the reaction time and the wetting state of the surface is very important. Single cell values correspond well between a forecasted and a dataset event. However, complex transport paths have a very high variability that is not reproducible with similar events. Further research is required to clarify if this is a result of the random walk approach or of the injection time of the particles.