Reservoir simulation is a CPU intensive task since it solves a set of fully implicit non-linear equations, the full-order model (FOM). Despite advances in computer processors, reducing simulation time is still a challenge especially for production optimization where hundreds of simulation runs are required. An acceptable approximation of reservoir dynamic model can be obtained when the model order is reduced. However, this leads to inaccuracies. These inaccuracies in the reduced-order models (ROMs) are addressed in here, using adaptive schemes to switch between the FOM and ROM. The proposed schemes are evaluated using coarse/fine grid homogenous and heterogenous case study models. These case studies are analyzed through 5-spot and inverted 5-spot patterns of production and injection wells for water flooding. Initially, the proper orthogonal decomposition (POD) is used to reduce the (FOM) for water flooding simulation. Snapshots are taken in FOMs’ runs to build ROMs. The average absolute deviation percent of the constructed ROMs relative to FOMs are less than 10% in most cases. The runtimes in ROMs are about 80% less than that of FOMs. A sensitivity analysis (based on case studies) is employed to determine the optimal period for snapshots and shock ranges. Three adaptive schemes (water breakthrough, injection-production intensity, and recovery factor) are proposed to switch between FOM and ROM in appropriate simulation timesteps. The best adaptive ROM scheme for heterogenous model is injection-production intensity while for homogenous reservoir model is recovery factor. This study shows increased accuracy for adaptive ROMs while simulation runtime is reduced by 50% for coarse homogenous and heterogonous reservoir models and 20% for the fine-grid model. ROM fidelity for inverted 5-spot pattern is more than that of 5-spot pattern.