Machine learning allows to create complex models if provided with enough data, hence challenging more traditional system identification methods. We compare the quality of neural networks and an ARX model when use in an model predictive control to command a drone in a simulated environment. The training of neural networks can be challenging when the data is scarce or datasets are unbalanced. We propose an adaptation of prioritized replay to system identification in order to mitigate these problems. We illustrate the advantages and limits of this training method on the control task of a simulated drone.