A modular approach to the design of neural networks for fault diagnosis in power systems

A modular approach to the design of neural networks for fault diagnosis in electrical networks of real size is described. Modularization is strictly based on functional criteria, rather than the usual topological criteria. This approach allows elimination of the problems inherent in this kind of application, which are large amounts of information to be processed, a high degree of uncertainty in the data, changes in the topological features, and sources of uncertainty. The most important characteristics of the model are the simplicity of the modules, the replicability of the training results, easy adaptation to topological changes, and high scalability. It allows for parallel implementation. A portion of a real distribution electrical network has been simulated