ON THE NONLINEAR PROPERTIES OF FEEDFORWARD NEURAL NETWORKS

In this paper, we present some analytical results on the nonlinear fea tures of the feedforward neural networks. The objective of our work wa s to further advance the understanding of the nonlinear characteristic s of the neural networks as the means of improving our ability in usin g the technology for developing power system applications. We show tha t the nonlinear characteristics in the output neurons act as a contrac ting mapping that maps the network output close to an average value de termined by the raw data scaling scheme. We show that the hidden neuro ns function as a set of switching elements that can smoothly move betw een linear, off and on states corresponding to different input pattern s. We also study the sensitivity of the neural network's output with r espect to input changes and provide a geometrical interpretation of th e hidden neuron nonlinear interactions. This analysis can be used as a basis in evaluating the performance of different feedforward neural n etworks. It can also be used as a basis in designing applications that fully utilize the potential of the technology in describing complex i nput/output relationships over a wide range.