Because most of the area in Westland is taken up by greenhouses, water hardly infiltrates into the ground which may result in flooding. Furthermore, rain showers are getting heavier, more variable and harder to predict due to climate change (Beersma et al., 2019). The space to find solutions for mitigating flooding is limited by the greenhouse area.

The waterboard of Delfland and the greenhouse horticulturists have found a small-scale solution that is called Rainlevelr. The solution focuses on releasing reservoir water before a predicted heavy rain event such that the reservoirs can capture precipitation during the actual event. The greenhouse horticulturists participate voluntarily in Rainlevelr. While both the horticulturists and the water board want to mitigate flooding, the greenhouse horticulturists also want to keep as much water as possible in their reservoirs for irrigation. When predicted rainfall differs from the actual rainfall, the resulting suboptimal strategy for releasing water from the reservoir can result in flooding the polder or wasting reservoir water. This raises the question of how to define an algorithm to assess trade-offs between the interests of different stakeholders.

A numerical model could predict polder water levels and reservoir water levels based on a weather forecast. A controller could then be added to the model to optimize the amount of water that should be released. The goal of this thesis is to design a Model Predictive Control (MPC) strategy for a polder system in Delfland by optimizing the valve setting of the Rainlevelr pipes in the reservoirs. With the proposed strategy the risks can be reduced, and possible trade-offs can be identified. Polder sections fail when the water level exceeds a threshold water level and reservoirs fail when their water level falls short.

The Python model POKKA was developed in order to study the effects of the valve settings on the water levels of both the reservoirs and the polder sections. The Kralingerpolder was used in a case study.

A MPC simulation showed that the controller was able to reduce the number of polder section threshold exceedances at the expense of reservoir threshold exceedances. However, the controller was not able to prevent flooding of the polder during the heavy rain event at the beginning of November 2023. The controller selected an optimal sequence of hourly valve settings out of a set of 100 realizations. Increasing this value to 1000 only leads to a small difference in predicted water levels whereas the computation time increased with a factor 10. A parallel computation over a period of a year took about 2 hours for 100 realizations. For this reason the remaining calculations were performed with 100 realizations. The controller implements the trade-offs by setting weights on the deviations of the polder water levels and the reservoir water levels from their reference level. A simulation showed that fewer polder sections fail if the weights favor of the polder sections. The controller computes the optimal valve settings based on weather predictions and its performance decreases with decreasing accuracy of the weather forecast. Its performance was improved by introducing a feedback loop where measured water levels were used to correct the model predictions. Increasing the Rainlevelr capacity has a large effect on the Model Predictive Control performance. The simulation showed that the controller came very close to completely preventing a level exceedance of the polder during the heavy rain event at the beginning of November if all reservoirs in the Kralingerpolder were part of the Rainlevelr initiative.