Integrated dynamic framework for predicting urban flooding and providing early warning

Urban flooding is a serious challenge in cities owing to global warming and rapid urbanization; thus, urban flood forecasting is required to reduce losses. Nevertheless, owing to the randomness and uncertainty of rainfall and ponding processes, providing accurate and stable ponding predictions using the existing prediction methods is difficult. To address these issues, this study proposes a time-varying criterion to improve the Bayesian model averaging (BMA) method and uses the time-varying BMA (TV-BMA) approach to develop an integrated model for predicting urban floods and providing early warning by dynamically coupling the results of the BMA and three machine-learning models. This integrated model was based on numerous measured data on rainfall and ponding processes, aiming to accurately predict real-time changes in ponding depth. The results show that the prediction accuracy of the TV-BMA model was 11.4–50.4 % higher than that of the BMA and individual models, and within the 95 % confidence interval. Moreover, the bandwidth of the TV-BMA model (0.038 m) was 19.1–53.1 % lower than that of the BMA (0.047 m) and individual models (0.063–0.081 m). This demonstrates that the TV-BMA model has significant advantages in correcting deviation and reducing uncertainty in the prediction results. Furthermore, the TV-BMA model has high accuracy in the early warning results for up to a 100-min forecast period (F-score > 0.8). The proposed TV-BMA model can predict urban floods with improved accuracy and stability. It may provide guidance for improving the existing urban flood prediction methods.