Numerical models of a water system are always based on assumptions and simplifications that may result in errors in the model predictions. Such errors can be reduced through the calibration of the model to in situ and/or satellite measurements of the system's state through the integration of models and data. Use of the Ensemble Kalman Filter (EnKF) with recent measurement data in operational forecast situations will significantly improve the success rate of the forecasts. The EnKF is a generic data assimilation method which has the advantage that its algorithm is relatively simple to implement and is also well-suited for highly non-linear models. An EnKF software module was implemented and applied successfully to SOBEK-River hindcasts for a section of the Rhine in 2002/ 2003 (El Serafy, 2003). In the present work, the EnKF was implemented as part of a generic data assimilation tool box and applied in combination with a Delft3D-Flow model of Osaka Bay. This test case provided more insight into EnKF performance with real-world operational forecasting systems for two and three dimensional flow regimes.
The primary driving force for the circulation in the Osaka Bay is tide, which is mainly diurnal, with spring range on the order of 2 m. In the northeastern section, five rivers discharge into Osaka Bay. Varying wind and river discharges drive the local salinity variations and leads to a locally salinity stratified three-dimensional circulation system. The aim of the application of EnKF is to improve the daily operational forecasts of salinity and current profiles for engineering activities in this stratified basin.
Due to operational constraints, a full EnKF was computationally too demanding, thus a simplification was chosen. This steady state Kalman filter (SSKF) was calibrated for the Osaka Bay model by assimilating hourly salinity and velocity components in two locations and four different vertical levels for the period 13-28 of Feb. 2002. The performance of the SSKF for improving the salinity and velocity components during the first 24 hours forecast is illustrated. It may be observed that using data assimilation, the model's predictions are better than those seen in the model without data assimilation, with the update effect disappearing over time. With newer measurements and new forecasts, this forecast window moves with time.
The present results show the practical feasibility of EnKF for data assimilation and are promising for application of EnKF to all sorts of other models used in hydrology and hydraulic engineering, for example water level, wave or run-off forecasting. Due to the simplicity and generic properties of the algorithm, these extensions can be realised with a limited amount of effort.
Co-funding of the present research by Kajima Technical Research Institute in Japan is gratefully acknowledged. |
