In this paper we study a non-minimum phase discrete time system with varying time-delay. We first propose several open loop control architectures based on non-linear neural networks and study their ability to handle the different difficulties of the control problem. All the methods are tested and compared to a baseline linear controller, on a simulated river system. This plant is submitted to perturbations corresponding to water withdrawals and lateral inflows. The above architectures are not able to cope with such perturbations. We then propose a model combining a feed-forward neural network based learning controller and a feedback adaptive controller. The performances of this model are compared to a similar architecture containing linear feed-forward and feedback controllers.