TABLE OF CONTENTS The new neural net techniques represented in the feasibility studies have to be validated in the future under real operating and time constraints in a control centre of an electric utility. This section discusses the procedure for training set generation and training execution for real world power systems either on a sequential computer or innovative dedicated neural net hardware. As an example we study the Kohonen network. However similar considerations apply to other types of neural network.
For the presented feasibility studies of the application of artificial neural nets to power system security assessment all network data were simulated with conventional load flow software. The power systems modelled are standard test systems. The load flow software as well as the Kohonen or the MLP simulator are currently running on a conventional UNIX workstation. The ANN simulators were specifically developed for this application. Earlier versions of the Kohonen simulator included a graphic interface which is currently transported to the X-Windows environment. The neural net simulator and the power system software communicate via files only. Furthermore, a software package for the statistical analysis of the weight vectors and the evaluation of the classification quality was developed. This software allows the evaluation of results produced by other Kohonen nets, whether simulated or implemented in hardware.
For the cases studied in this work, the performance of the Kohonen simulator proved to be sufficient. Training times...
