This paper presents the application of Multilayer Perceptron (MLP) and Radial Basis Function (RBF) neural network for managing active and reactive powers of distributed generation (DG) units in distribution systems. A two-stage intelligent technique is proposed using an iterative interior-point algorithm optimization procedure for collecting the optimal power settings of several DG units in the first stage. In the second-stage, the optimal data obtained from the optimization process are then used for training the MLP and RBF neural networks which will then predict the next time step of active and reactive power references of each DG unit for online application. The results show that the MLP network has the ability in predicting the optimal power reference of the DG units with small errors compared to the RBF network. However, the RBF network converges faster compared to the MLP network.