Adaptive neuro-fuzzy modeling of electric arc furnaces

This thesis investigates the use of Adaptive Neuro-Fuzzy Networks (ANFN) as means of implementing algorithms suitable for Nonlinear Black-Box Modeling and Prediction. One area of particular importance is the design of networks capable of modeling and prediction of the behavior of systems that involve complex, multi-variable processes with time-variant parameters. This thesis shows that ANFNs lend themselves well to nonlinear black-box modeling/prediction. To illustrate the applicability of the neuro-fuzzy networks, a case study involving electric arc furnaces (EAF) is presented here. EAF represents a complex, nonlinear load with stochastic behavior and as such prediction/Modeling of EAF characteristics is a demanding task. This thesis first describes the notion of nonlinear black-box modeling and shows that neuro-fuzzy networks are model structures suitable for nonlinear black-box prediction. A uniform treatment of artificial neural networks (ANN), fuzzy logic systems (FLS) and ANFNs are presented in which applications of these methodologies are explored to model v-i characteristic of an operational EAF whose measured data are used for training and validation of all networks. The successful developments of EAF models, using these different network architectures has been described. This thesis also presents application of two classes of hybrid neuro-fuzzy networks, namely, feedforward and recurrent, to the solution of a particular complex problem that is long-term prediction. A frame of reference is built, based on which other nonlinear black-box techniques suitable for single-step/multi-step prediction can be evaluated. Successful implementations of feedforward/recurrent neuro-fuzzy predictors are described and their performances are illustrated using the results obtained from adaptive neuro-fuzzy networks and recorded data.