Machine Learning-Based Early Warning System for Urban Flood Management

Abstract

With the growth in urban population and other pressures, such as climate change, the impact and severity of urban flood events are likely to continue to increase. “Intelligent water networks” are viewed as the way forward to ensure that infrastructure services are flexible, safe, reliable and economical. Reduction of flood-risk from urban drainage and sewerage infrastructure is likely to require increasingly sophisticated computational techniques to keep pace with the level of data that is collected both from meteorological and online water monitoring systems in the field. This paper describes and characterises an example of an Early Warning System (EWS), designated "RAPIDS" (RAdar Pluvial flooding Identification for Drainage System) that deals with urban drainage systems and the utilisation of rainfall data concurrently to predict flooding of multiple urban areas in near real-time using a single multi-output Artificial Neural Network (ANN). The system has the potential to provide early warning for decision makers within reasonable time, this being a key requirement determining the operational usefulness of such systems. Computational methods that require hours or days to run will not be able to keep pace with fast-changing situations such as manhole flooding or Combined Sewer Overflow (CSO) spills and thus the system developed is able to react in close to real time. This paper includes a sensitivity analysis and demonstrates that the - predictive capability of such a system based on actual rainfall is limited to a maximum of the Time of Concentration (ToC) of each node being modelled. To achieve operationally useful prediction times, predictions of rainfall as input signals are likely to be needed for most urban drainage networks.