This article presents an efficient solution for predictive control based on neural networks with feedfoward multilayer, as a model for a thermoelectric module. It is shown the capability of a neural network to learn the entire nonlinear dynamics and the advantage of using these nonlinear models for the calculation of the predicted variables. It is also suggested a new control law capable of minimize the cost function using the Newton-Raphson and the descendent gradient optimization rules. For this application it is shown that a significant reduction in the number of iterations and application in real-time systems when compared to other optimization techniques.