Polder areas as typical in the Netherlands require real time control on the water levels. Efficient pumping is a subject with increasing interest due to a required use of 20% green energy by 2020 set by the government. In 2010, the energy requirements already exceeded 175 GWh per year. For optimal control, water level predictions are important. However, they are uncertain due to errors in model structure, parameter estimation, initial states and observed forcing data. Especially the rainfall was found to contribute to water level prediction uncertainty. The uncorrected radar rainfall forcing data available in real time underestimate the corrected radar rainfall measurement over 100%, worsening with increasing rainfall intensities.

The aim of the research was to asses whether data assimilation can reduce the uncertainty on the water level predictions in a Delfland polder. The polder is modelled with a conceptual rainfall-runoff model in Sobek RR, while the data assimilation is implemented as an ensemble Kalman filter in OpenDA.
In order to manipulate the states in Sobek RR for data assimilation, the model was extended by a black-box wrapper that was newly created in Java for this research.

To asses the theoretical applicability of data assimilation in a polder area, four twin experiments were carried out, which differed only in the model components to which noise was applied. The four set-ups included the use of a rainfall multiplier, additive noise on the model states, additive noise on the groundwater state and a single multiplier for all states. The results were compared visually and objectively through the probabilistic Nash Sutcliffe, a model efficiency coefficient.

All set ups for the twin experiment generated good results for the assimilated water level, mimicking the observed water levels closely with probabilistic Nash Sutcliffe values ranging between 0.92 to 0.97. However, only the set up with the rainfall multiplier exhibited good score measures for the hidden states, with an average of 0.6, showing a significant improvement over the case without assimilation. The other twin experiments failed to represent the hidden states, with average score measures below zero.

The data assimilation set-up with the rainfall multiplier was then applied using water level observations from Tedingerbroek polder. The pump pattern and magnitude of water levels of the modelled and observed water level differed significantly. The influence of the different model components on the modelled water level was assessed, from which was concluded that these could not account for the difference in modelled and observed values. Possible explanations are due to a non-representative measurement point in the polder or additional unknown water fluxes coming into the polder.

Recommended further research should focus on creating a larger observation network; implementing multiple water level observation locations, collecting pump discharges and taking measurements of the storage in greenhouse basins. The last could give a method to validate the assimilation of the hidden states, whereas the pump discharges should be used as additional observation in the data assimilation.