Real time learning control of an emergency turbo-generator plant using structurally adaptive neural networks

This paper describes the real-time application of a novel nonlinear control architecture, the "rectangular local linear controller" (RLLC) network, to a tracking problem for a highly nonlinear plant: an emergency turbo-generator. The nonlinear controller is composed of a weighted combination of a number of linear controllers, each of which is locally activated in some operating region of the plant. Each local controller is designed herein using the adaptive "linear quadratic regulator" (LQR) approach. Furthermore, the "rectangular local linear model" (RLLM) network is used to model the plant, thus providing means for obtaining a corresponding linear model for each local controller needed by the design. The structure and the parameters of the RLLM network are automatically adapted on-line using the "on-line adaptive k-tree lattice learning" (ONALAL) algorithm, leading to a parsimonious model. Experimental results show the practical viability of the new proposed approach.