Data: 17/09/2014 à 19/09/2014
Local: São Paulo - Brazil
Application of Unscented Kalman Filter as Data Assimilation Method to Flood Prediction System of River Yoshii (PAP014405)
Flood forecasting and early warning systems
The purpose of this study is to estimate the applicability of unscented Kalman filter, which is one of methods to optimize the state valuable in governing equations in order to fit the temporal water surface level from past time to current time. Unscented Kalman filter is a method to optimize the state valuable with using static two order valuable of the observation and the system noise. However particle filter, which is based on the Monte method, is effective to solve the optimization in inverse problem. In this study, we improved the accuracy of water surface level forecast system, which predicts 40km reach from river mouth in River Yoshii, Okayama prefecture in Japan. One dimensional hydrodynamics model was adopted as system equations. Water surface level at downstream and discharge at upstream and downstream were chosen as state variable, which were parameters in governing equations. These state variables were optimized by Unscented Kalman filter in order to fit water surface level at current time. Furthermore predict computation is performed with optimized state variable as the initial condition. The temporally discharges by solving with distributed runoff model was given to the boundary conditions at upstream was given. On the other hands, predicted water surface level before 6 hours was given to the boundary conditions at downstream. We estimated the accuracy of water surface level at 1, 3 and 6 hours after by comparing the predicted water surface level with the observed one. We judged the unscented Kalman filter to be proper as the technique that a flood prediction was available to. Moreover this data assimilation methods can suppose temporal discharges. As a result of having compared the conversion discharges for the observated water level with the estimated discharges, the error understood a certain around 10% of ranges. This technique was shown to be validity from compatibility to observation water level as the technique that was applicable to a flood prediction system.