Model uncertainty reduction using data assimilation in the Dutch polder area of Delfland
By Petra Izeboud
Many contributors to model errors exist, including uncertainty in forcing data, output data and model structure. Errors in precipitation measurement can be huge, causing wrong simulated outflow, but also wrong initital states for the next simulation step. In its turn, wrong initial states can cause large deviations in the predicted runoff.
In the canal drainage system of Delfland, a rainfall runoff model is used to predict the discharge of the polder to the boezem via pumps. This discharge is used as input to a model predictive control system that determines the optimal course of action for the larger pumps in the boezem.
In order to improve the input for the model predictive control, the initial states for each model run can be updated using the measured discharge to the boezem system. Another method is to propagate the expected error in the rainfall measurements using processed raindata that becomes available upto 38 hours after the rainfall event.
The aim of this thesis is to see how the prediction of the discharge in the rainfall runoff model can be improved using data assimilation. State updating using discharge measurements and Kalman filtering is compared to error prediction of the rainfall using processed rainfall data and the data that is available in real time. Of interest is how well these methods can reduce the uncertainty due to rainfall or the model uncertainty.