Runoff Prediction in Small Rivers Using Dynamic Parameter Fitting with Observed Rainfall Data

Floods in a small river basin measuring a few square kilometers can sometimes cause the water levels to rise more than 1 m within 10 min. Therefore, the importance of runoff prediction within tens of minutes is greater in small rivers than in large ones. In this study, we developed a runoff prediction method that uses only observed rain by considering the lag time, representing the time from rainfall to runoff, as prediction lead time. The developed prediction method (DSG) employed the Storage Function model for the runoff simulation and a Genetic Algorithm (GA) for dynamically automatic parameter fitting that was performed in each time step of a rain event. The DSG was applied to six heavy rain events that caused flooding in the Kohatsu River in Japan from 2020 to 2022. The simulation results showed that the GA could effectively determine the optimal parameters to ensure that past simulation depths agreed with the observed ones at each time step of the simulation. Prediction results of the DSG were compared with those of the Storage Function Model with static pre-calibrated parameters (SSG), the simple persistence model (SP), and the gradient persistence model (GP) used as benchmarks. The prediction results of the developed model had the slightest total error of the water level hydrograph (Average $(\mathrm{TE})=0.17$) compared to the predictions of the SSG, SP, and GP models across all six heavy rain events. In addition, the coefficient of determination $(R^2)$ for the flow depth prediction using the developed method (DSG) was higher than 0.9 for each rain event.