We propose an end-to-end trained neural networkarchitecture to robustly\npredict the complex dynamics of fluid flows with high temporal stability. We\nfocus on single-phase smoke simulations in 2D and 3D based on the\nincompressible Navier-Stokes (NS) equations, which are relevant for a wide\nrange of practical problems. To achieve stable predictions for long-term flow\nsequences, a convolutional neural network (CNN) is trained for spatial\ncompression in combination with a temporal prediction network that consists of\nstacked Long Short-Term Memory (LSTM) layers. Our core contribution is a novel\nlatent space subdivision (LSS) to separate the respective input quantities into\nindividual parts of the encoded latent space domain. This allows to\ndistinctively alter the encoded quantities without interfering with the\nremaining latent space values and hence maximizes external control. By\nselectively overwriting parts of the predicted latent space points, our\nproposed method is capable to robustly predict long-term sequences of complex\nphysics problems. In addition, we highlight the benefits of a recurrent\ntraining on the latent space creation, which is performed by the spatial\ncompression network.\n