We present a two-level strategy to improve robustness against uncertainty and model errors in life-cycle flooding optimization. At the upper level, a physics-based large-scale reservoir model is used to determine optimal life-cycle injection and production profiles. At the lower level these profiles are considered as set points (reference values) for a tracking control algorithm, also known as a model predictive controller (MPC), to optimize the production variables over a short moving horizon based on a simple data-driven model. We used a conventional reservoir simulator with gradient-based optimization functionality to perform the life-cycle optimization. Next, we applied this optimal strategy to a set of reservoir models with markedly different geological characteristics. We compared the performance (oil recovery) of these models when applying the life-cycle strategy with and without the corrections provided by the data-driven algorithm and the tracking controller. In this theoretical study we observed that the use of the lower-level controller enabled successful tracking of the reference values provided by the upper-level optimizer. In our example, a performance drop of 6.4 % in net present value, caused by differences between the reservoir model used for life-cycle optimization and the (synthetic) true reservoir, was successfully reduced to only 0.5% when applying the two-level strategy. Several studies have demonstrated that model-based life-cycle production optimization has a large scope to improve long-term economic performance of water flooding projects. However, because of uncertainties in geology, economics and operational decisions, such life-cycle strategies cannot simply be applied in reality. Our two-level approach offers a solution to realize the theoretical potential of life-cycle optimization in a more operational setting.