This paper revisits system identification and shows how new paradigms from machine learning can be used to improve it in the case of non-linear systems modeling from noisy and unbalanced dataset. We show that using importance sampling schemes in system identification can provide a significant performance boost in modeling, which is helpful to a predictive controller. The performance of the approach is first evaluated on simulated data of a Unmanned Surface Vehicle (USV). Our approach consistently outperforms baseline approaches on this dataset. Moreover we demonstrate the benefits of this identification methodology in a control setting. We use the model of the Unmanned Surface Vehicle (USV) in a Model Predictive Path Integral (MPPI) controller to perform a track following task. We discuss the influence of the controller parameters and show that the prioritized model outperform standard methods. Finally, we apply the Model Predictive Path Integral (MPPI) on a real system using the know-how developed here.