Model Predictive Control (MPC) is an excellent control strategy that has high performance and a great ability to deal with multivariate process interactions; constraints on both system inputs and states; and real-time optimization requirements. However, some control problem drawbacks such as process non-linearity, or the non-convexity of the resulting optimization problem generate a higher computational cost for real-time MPC implementation, requiring embedded devices with a higher memory and processing capacity. Consequently, MPC is mostly used in processes with large time constants and/or where devices with high computational performance are available. In this article a controller based on a Neural Network trained from the data generated by a suitable MPC is presented. The proposed controller uses a Recurrent Neural Network to accurately predict the control input based on the previous training data, and once trained the RNN replaces the MPC completely. This reduces the computational cost by not requiring to solve the optimization problem online. The effectiveness of the proposed approach is demonstrated through simulations on a multivariate four coupled-tanks system.