Steam Temperature Regulation in Fossil Power Plants using Neural - Adaptive Strategies

In this report, neural-adaptive controllers are described for steam temperature regulation in fossil fuel power plants. The algorithms were developed in response to the need for improving the steam regulation properties of the existing controllers. These controllers were observed to exhibit temperature spikes in the event of changing operating conditions. Since the PI controllers have no auto-tuning and operate based on linear principles, these controllers cannot adequately accommodate nonlinear and time-dependent system behavior. Steam temperature behavior in power plants is strongly dependent on load output of the boiler, the cleanliness, and the main steam flow and fuel properties. Moreover, the behavior changes as a result of equipment replacement or malfunctions. The heat transfer from the flue gas to steam makes this system inherently nonlinear. In this report, a new class of adaptive controllers was developed for steam temperature regulation. The controllers have the ability to auto-tune their parameters on-line and respond to nonlinear dynamic characteristics in the system. The specific problem that we address in this report concerns temperature regulation at the outlet of the reheater or superheater. Two different controllers are developed for this purpose. The first is a linear neural adaptive controller that consists of a pole-placement structure and an update law for adjusting linear control parameters. The controller also accommodates plant saturation effects by including them in the adaptation law and effects of set-point disturbances by including integral action. The second controller is a nonlinear controller that not only includes above linear part but also includes a feedback linearizing substructure together with a TANN (Theta (0) Adaptive Neural Networks) that allows the adjustment of nonlinear parameters. Both controllers are developed using system-identification model of the reheater dynamics. The controllers are implemented on a full-scale and detailed simulator at the EPRI Instrumentation & Control Center, Kingston, Tennessee using the Foxboro DCS (Distributed Control System). The performance of the controllers is compared to the present PI implementation. An order of magnitude improvement in the settling time after a major temperature upset was achieved as well as an order of magnitude decrease in the maximum temperature deviation. The performance improvement with the neural controllers was observed both in the context of disturbance rejection and command following. Thesis Supervisor: Anuradha M. Annaswamy Title: Principal Research Scientist