Conference Agenda

In the Ruhr region of Germany, pumping stations are often used in urban areas to drain water from rivers due to land subsidence caused by coal mining. In case of pump failure, these areas face an immense threat to human life and property due to the short warning time of flooding. Data-driven models can predict flood inundation in real time, allowing for the timely initiation of protective measures. In this study, a new approach based on a data-driven model is developed using a convolutional neural network (CNN) to dynamically predict the maximum water depth of the next 24 h for fluvial and pluvial flooding in real time. The model is trained with physically based pre-simulated scenarios considering rainfall and runoff curves as inputs. The resulting forecasting system provides accurate real-time predictions of flood extent and water depth for different pump failure scenarios up to an average root mean square error of 0.038 m and an average critical success index of 0.953. The developed forecasting system is a suitable approach for operational control systems that allow accurate prediction in real time. A further research issue represents the dynamic forecast by the developed approach, predicting the current water depth over an event.