In this research, 18 artificial neural network (ANN) architectures are implemented, using two optimization algorithms: gradient descent and Levenberg Marquardt, to predict the wind direction as the input of a wind turbine direction control system, these optimization methods have been chosen because they allow convergence towards the minimum value of the function. To feed the neural network is used a real data set composed of six input variables, which allows us to reach a validation root mean square error RMSE of 1.63E-7%. A comparison between the different architectures has been done through the architecture's RMSE comparison, where it was observed that the best architecture for the network trained with the training algorithm through gradient descent was [7 5 1] with 0.0021 of MSE, 0.0013 of RMSE and for the artificial neural network with the training algorithm of Levenberg-Marquardt the best architecture was [7 5 3 1] with 8.73E-19 from MSE and 1.63E-7 from RMSE.