Self-organized neuro-fuzzy identifier with automatic rule generation for a fossil-fuel power plant control

In this thesis, a new paradigm consisting of integration of Fuzzy Logic systems into multi-layer feedforward neural networks is described. The neuro-fuzzy network developed in this research work utilizes a novel approach to integrate the fuzzy rule set within the network weighting matrices without losing the linguistic characteristic of the fuzzy reasoning. The fuzzy rules are resolved automatically by Genetic Algorithm (GA) training method. The GA establishes fuzzy rules by minimizing errors between fuzzy system outputs and plant's response. A unique approach is developed to encode the fuzzy rules in compound chromosomes of GA. The GA process operated as multiple processes in compound chromosomes. A learning algorithm based on Error Back-Propagation is formulated from the network structure in order to modify the width and mean of the membership functions. The achieved system is an intelligent self-organized neuro-fuzzy system. The developed neuro-fuzzy structure is utilized in modeling the dynamics of a boiler/turbine. This rule-based identifier is applied in a predictive controller of the power unit. The Neuro-fuzzy identifier provides the prediction of plant outputs to be used in an optimization algorithm. Evolutionary programming (EP) is chosen as the computational optimization procedure for this task. The obtained system is a novel knowledge-based predictive controller using EP as an optimizer. In addition, a new defuzzification method is introduced for fuzzy reasoning. This method approximates the Center of Gravity defuzzification method with the first of maxima in output membership functions.