摘要：Pluvial floods in urban areas are caused by local, fast storm events with very high rainfall rates, which lead to inundation of streets and buildings before the storm water reaches a watercourse. An increase in frequency and intensity of heavy rainfall events and an ongoing urbanization may further increase the risk of pluvial flooding in many urban areas. Currently, warnings for pluvial floods are mostly limited to information on rainfall intensities and durations over larger areas, which is often not detailed enough to effectively protect people and goods. We present a proof‐of‐concept for an impact‐based forecasting system for pluvial floods. Using a model chain consisting of a rainfall forecast, an inundation, a contaminant transport and a damage model, we are able to provide predictions for the expected rainfall, the inundated areas, spreading of potential contamination and the expected damage to residential buildings. We use a neural network‐based inundation model, which significantly reduces the computation time of the model chain. To demonstrate the feasibility, we perform a hindcast of a recent pluvial flood event in an urban area in Germany. The required spatio‐temporal accuracy of rainfall forecasts is still a major challenge, but our results show that reliable impact‐based warnings can be forecasts are available up to 5 min before the peak of an extreme rainfall event. Based on our results, we discuss how the outputs of the impact‐based forecast could be used to disseminate impact‐based early warnings. Plain Language Abstract Pluvial floods are caused by local rain storms with extreme rainfall rates, which leads to immediate flooding of streets and buildings in urban areas. These events are expected to increase in the future due to climate change and growing urban areas. Pluvial floods are directly caused by a rainstorm, which gives citizens and emergency responders usually only a few minutes to act. Existing forecasting systems for pluvial floods are limited to rainfall forecasts that neither provide information about where a flood might occur nor how severe the impacts will be. Here, the main challenge is that current computer models that predict inundation take too long to run to release flood forecasts early enough. We present a new inundation model that can predict inundation for an upcoming flood event in a fraction of the time of existing models. We combine this model with models that predict the spreading of contamination (e.g., from a car accident) and the damage to residential buildings. For a real flood event we can show that this information can be released up to 5 min before the rainfall peak, which gives citizens and emergency responders the opportunities to safe lives and protect important valuables.
