Um novo conceito de inteligência artificial baseado nas redes da lógica paraconsistente anotada de dois-valores para tratamento de incertezas na estimativa de seção em falta do sistema elétrico de potência

Abstract

This thesis proposes a novel concept of artificial intelligence based on a contemporary and non-classical logic called Paraconsistent Annotated Logic of two-values (PAL2v) applied on fault section estimation in an electric power system. The fault section estimation is a decision-making problem because, under abnormal electrical operating conditions, a large volume of alarms are generated in a short time. It is up to the control center operator to make the most appropriate decision to isolate the fault section. In this context, the importance of this work is to develop a methodology that aids the operator in decision making in stressful situations. In addition, considering that there are no reports in the literature of LPA2v for fault section estimation, the work contributes to the innovative aspect. Its innovation lies on the problem´s approach, since unlike conventional methodologies that establish binary solutions (fault section is 1 and not fault is 0), LPA2v admits uncertainties in its own decision making. Likewise, it is possible to ensure higher reliability in the solutions passed to the operator, because in the presence of an uncertain solution, it avoids hasty decisions in binary 1. LPA2v is evidence-based, and the more evidence it obtains for a fault scenario, the more equitable the diagnoses/solutions will be. To extract the evidence, three heuristic functions were used that employ inference rules using reported alarms from Supervisory Control and Data Acquisition (SCADA). The use of three functions avoids dependence on probabilistic or empirical values. Furthermore, two paraconsistent networks were developed, both extracting evidence from the heuristic functions. The first paraconsistent analysis network call with LPA2v was tested on a 138/203 kV subsystem in southern Brazil and its results were compared with exact mathematical models to solution optimizations. The second analytical paraconsistent artificial neural network call was tested in a 345 Kv transmission system and was compared with three propositional logic models that incorporate fuzzy interval values and spiking neural network. Finally, both paraconsistent networks showed: 1- robustness in the comparisons of results, because in cases of false and failed alarms, they detected uncertain solutions as expected; 2- easy implementation in different electrical systems, as it does not require training and elaborate construction of rules or pattern; 3- intuitiveness in estimating faults in the SEP, as it goes beyond methods that offer conventional solutions (0 or 1) for the system operator.