Various rainfall-runoff models have been developed and used to predict floods in watersheds. Conceptual rainfall-runoff models have been widely used in practice due to their high reliability and applicability. However, these models simplify the rainfa...
Various rainfall-runoff models have been developed and used to predict floods in watersheds. Conceptual rainfall-runoff models have been widely used in practice due to their high reliability and applicability. However, these models simplify the rainfall-runoff process, making it difficult to incorporate the physical processes into runoff predictions. Additionally, the models have various parameter uncertainties, requiring calibration using a sufficient amount of observational data. On the other hand, physically-based rainfall-runoff models physically consider the flow within the watershed, providing relatively accurate results. However, their high computational cost and numerical instability limit their practical application.
In this study, we developed a flood prediction model that integrates a physically-based dynamic wave model, an instantaneous unit hydrograph (IUH) model, and a deep learning model to achieve both accuracy and efficiency. We derived and interpolated the IUH based on the dynamic wave simulation and the power-law relationship between peak characteristics of IUH and rainfall excess intensity. In addition, we estimated rainfall losses using the Green-Ampt model and the NRCS-CN model. To estimate the parameters of each rainfall loss model, we incorporated real-time GLDAS-Noah land surface modeling data and rainfall time series data into the LSTM model, a deep learning model. Finally, we predicted the flood hydrograph of the watershed through convolutional integration.
We applied our flood prediction model to both ideal and real watersheds. The LSTM model for rainfall loss estimation predicted the effective rainfall with high accuracy, achieving a Nash-Sutcliffe efficiency and correlation coefficient of over 0.9 and a normalized mean absolute error and normalized root mean square error below 0.15. In the flood prediction results, we obtained high accuracy with Nash-Sutcliffe efficiency between 0.55 and 0.9 and coefficients of determination between 0.67 and 0.95.
Our physically-based IUH can be derived using only the topography and Manning's roughness coefficient of the watershed, without the need for parameter calibration using observed data. This makes it advantageous for application to ungauged watersheds or watersheds with limited observational data. Additionally, the derived IUHs account for nonlinear rainfall-runoff relationship, as it is used as a function of rainfall excess intensity for a single watershed. In addition, we efficiently interpolated the IUH for arbitrary rainfall excess intensity using a power-law-based technique. Furthermore, the derivation of the flood hydrograph through convolutional integration enables instantaneous and efficient flood prediction.
By incorporating deep learning models and real-time watershed data for rainfall loss estimation, our flood prediction technique can reflect the moisture conditions of the watershed during flood events and consider various influencing factors related to rainfall loss. This approach reduces uncertainties in parameter estimation for rainfall losses and minimizes the need for subjective judgments.
The flood prediction model in this study is expected to provide several advantages for flood damage prevention and the operation of water control structures. Except for the initial construction processes of optimizing LSTM and deriving IUHs, flood prediction can be completed rapidly within minutes, making it highly efficient. Moreover, by considering real-time moisture conditions in the watershed and physically reflecting the rainfall-runoff process, the model ensures the accuracy of flood prediction.