Correcting mean areal precipitation forecasts to improve urban flooding predictions by using long short-term memory network

Urban flooding is a critical challenge in metropolitan cities around the world; thus, urban flood forecasting is required to support water-related managers in mitigating damage. Nevertheless, the accuracy of rainfall forecasting systems remains limited; for example, the predictions of radar-based systems are often inaccurate for heavy rainfall events. This study proposes a framework that couples a forecasting system and a developed 1D/2D urban hydrological model to predict water levels and inundation phenomena in an urban catchment. In the framework, a long short-term memory (LSTM) network uses the quantitative precipitation forecasts (QPFs) of the McGill Algorithm for Precipitation nowcasting by Lagrangian Extrapolation (MAPLE) system to reproduce three-hour mean areal precipitation (MAP) forecasts. A coupled 1D/2D urban hydrological model was also developed in this study. The Gangnam urban catchment located in Seoul, South Korea, was selected as a case study for the proposed framework. To train and test the LSTM model, a database was established based on 24 heavy rainfall events, 22 grid points from the MAPLE system and the observed MAP values estimated from five ground rain gauges. The corrected MAP forecasts were input into the developed coupled model to predict water levels and relevant inundation areas. The results indicate the viability of the proposed framework for generating three-hour MAP forecasts and urban flooding predictions. This study demonstrates that despite slightly underestimating extreme values of rainfall and peak water levels for certain events, the framework has high practicability and can be used to improve MAP forecasts and urban inundation forecasts.