الفهرس 
GENERAL REVIEW ON ARTIFICIAL NEURALOnly 14 pages are availabe for public view
 |  | from | 140 |  |  |
| | | from | 140 | | |
Abstract
Automatic power stabilization control is the desired objective
for any reactor operation, especially, nuclear power plants. A major
problem in this area is inevitable gap between a real plant and the
theory of conventional analysis and the synthesis of linear time
invariant systems. In particular, the trajectory tracking control of a
nonlinear plant is a class of problems in which the classical linear
transfer function methods break down because no transfer function
can represent the system over the entire operating region. There is
a considerable amount of research on the model-inverse approach
using feedback linearization technique. However, this method
requires a prices plant model to implement the exact linearizing
feedback. For nuclear reactor systems, this approach is not an
easy task because of the uncertainty in the plant parameters and
un-measurable state variables. Therefore, artificial neural network
(ANN) is used either in self-tuning control or in improving the
conventional rule-based expert system.
The main objective of thesis to suggest an ANN, based selflearning
controller structure. This method is capable of on-line
reinforcement learning and control for a nuclear reactor with a
totally unknown dynamics model. Previously, researches are
based on back-propagation algorithm. Back-propagation (BP), fast
back-propagation (FBP), and Levenberg-Marquardt (LM),
algorithms are discussed and compared for reinforcement
learning. It is found that, LM algorithm is quite superior.
Three Mile Island (TMI), pressurized water reactor (PWR) is
selected as nuclear power plant. The TMI model was developed,
as well as proposed NN-controller model and the conventional
proportional-integral derivative (PIO), controller model. All models
are simulated under MATLAB/SIMULINK package. The method is
based on a neural network model that embodies the nonlinear
behavior of nuclear reactors and an iterative method to determine
the one-step-ahead predictive control input. TMI modeling results
either response to reference power variation or response to
external reactivity variation concludes that, NN-controller using LM
algorithm is the superior to conventional PIO controller. This
means that, the suggested NN-controller has excellent robustness
and performance features