For the prediction of river floods, WL / Delft Hydraulics uses numerical models, known as Flood Early Warning Systems (FEWS). Rainfall-runoff models (RR-models) are important components in FEWS. RR-models often contain large uncertainties, which are due to model imperfections such as:
- Model errors, for example due to a lack of physical knowledge, or simplifications in the model to avoid high computation time.
- Errors in the input data of the model, i.e. uncertainties in predicted or observed rainfall or evapotranspiration. Data assimilation is an efficient and often-applied method used to reduce uncertainties in models and thus improve the model's performance. Data assimilation means that models and observations are integrated in some (optimal) way. The Kalman filter (KF) is a well-known example of a data assimilation technique that is often used in practise. The standard KF was developed for linear dynamic systems. However, in practise most physical systems and models are not linear. In order to apply Kalman filtering to these models, the algorithm must be appropriately adjusted, extended or approximated. The ensemble Kalman filter (EnKF) may be mentioned as one such alternative algorithm. One of the main advantages of the EnKF is that it is a generic method, i.e. it can be applied to virtually any dynamic model in discrete time.
The feasibility of the EnKF for input correction in RR-models was investigated within the scope of this project. The EnKF was applied to several RR-models, varying from very simple approaches to models with a more realistic representation of the physics. The final realistic case dealt with an application of the EnKF to the HYMOD RR-model. A sensitivity analysis was performed by varying the parameters in the model to assess the best statistics for the representation of the uncertainties in the rainfall and evapotranspiration. The results of this research indicate that the EnKF offers a valuable method for input correction in RR-models.
This project was executed in cooperation with Technical University Delft (MSc study). |
