A Rapid Forecast Method for the Process of Flash Flood Based on Hydrodynamic Model and KNN Algorithm

Using hydrodynamic models to carry out early warning and flash floods forecasting is an essential measure for loss reduction. Nevertheless, many current hydrodynamic models lack the necessary forecasting timeliness. To address this limitation, a method combining a hydrodynamic model with the K nearest neighbours (KNN) algorithm is proposed to facilitate the rapid prediction of flash flood processes. With the rainfall sequence as the input data and the simulation results of the hydrodynamic model as the target data, the rapid forecast of water depth, water velocity and discharge are achieved. Then the Baogai Temple basin is utilized as a case study, and the rapid forecast model (RFM) is established and subjected to verification for reliability and timeliness. The results demonstrate that the established model exhibits remarkable accuracy, with 99% of the test data effectively limiting the error of accumulated inundation extent within 20%. Furthermore, the Nash-Sutcliffe efficiency (NSE) for cross-sectional discharge achieves a value of 0.98. In 75% of rainfall scenarios, both the maximum average water depth and velocity errors for the cross-sections are effectively confined to 7.5% and 10%, respectively. The model also boasts a substantial improvement in computational efficiency, enabling it to complete the prediction of the flooding process for the next 10 h within 25s. This enhancement offers valuable lead time for emergency decision-making and highlights its extensive application potential in managing flash floods.