Urban flood forecasting using a hybrid modeling approach based on a deep learning technique

Abstract

Climate change is contributing to the increasing frequency and severity of flooding worldwide. Therefore, forecasting and preparing for floods while considering extreme climate conditions are essential for decision-makers to prevent and manage disasters. Although recent studies have demonstrated the potential of long short-term memory (LSTM) models for forecasting rainfall-related runoff, there remains room for improvement due to the lack of observational data. In this study, we developed a flood forecasting model based on a hybrid modeling approach that combined a rainfall-runoff model and a deep learning model. Furthermore, we proposed a method for forecasting flooding time using several representative rainfall variables. The study focused on urban river basins, combined rainfall amounts, duration, and time distribution to create virtual rainfall scenarios. Additionally, the simulated results of the rainfall-runoff model were used as input data to forecast flooding time under extreme and other rainfall conditions. The prediction results achieved high accuracy with a correlation coefficient of >0.9 and a Nash[ndash]Sutcliffe efficiency of >0.8. These results indicated that the proposed method would enable reasonable forecasting of flood occurrences and their timing using only forecasted rainfall information. HIGHLIGHTS A flood forecasting model based on hybrid modeling that combines a R–R model and a LSTM model is developed, and a method for forecasting floods using representative rainfall variables is proposed.; This study combined rainfall amount, duration, and distribution to create virtual rainfall scenarios.; The simulated results of the R–R model were used as input data to forecast flooding time under various rainfall conditions.;

Keywords

• flood forecasting
• flooding time
• long short-term memory neural network
• storm water management model
• urban stream