DESIGN AND IMPLEMENTATION OF INTELLIGENT MONITORING SYSTEMS FOR THERMAL POWER PLANT BOILER TRIPS

Steam boilers represent the main equipment in the power plant. Some boiler trips may
lead to an entire shutdown of the plant, which is economically burdensome. An early
detection and diagnosis of the boiler trips is crucial to maintain normal and safe
operational conditions of the plant. Numbers of methodologies have been proposed in
the literature for fault diagnosis of power plants. However, rapid deployment of these
methodologies is difficult to be achieved due to certain inherent limitations such as
system inability to learn or a dynamically improve the system performance and the
brittleness of the system beyond its domain of expertise. As a potential solution to
these problems, two artificial intelligent monitoring systems specialized in boiler trips
have been proposed and coded within the MA TLAB environment in the present work.
The training and validation of the two systems have been performed using real
operational data which was captured from the plant integrated acquisition system of
JANAMANJUNG coal-fired power plant. An integrated plant data preparation
framework for seven boiler trips with related operational variables, has been proposed
for the training and validation of the proposed artificial intelligent systems. The feedforward
neural network methodology has been adopted as a major computational
intelligent tool in both systems. The root mean square error has been widely used as a
performance indicator of the proposed systems. The first intelligent monitoring
system represents the use of the pure artificial neural network system for boiler trip
detection. The final architecture for this system has been explored after investigation
of various main neural network topology combinations which include one and two
hidden layers, one to ten neurons for each hidden layer, three types of activation
function, and four types of multidimensional minimization training algorithms. It has
been found that there was no general neural network topology combination that can
be applied for all boiler trips. All seven boiler trips under consideration had been
detected by the proposed systems before or at the same time as the plant control system. The second intelligent monitoring system represents mergmg of genetic
algorithms and artificial neural networks as a hybrid intelligent system. For this
hybrid intelligent system, the selection of appropriate variables from hundreds of
boiler operation variables with optimal neural network topology combinations to
monitor boiler trips was a major concern. The encoding and optimization process
using genetic algorithms has been applied successfully. A slightly lower root mean
square error was observed in the second system which reveals that the hybrid
intelligent system performed better than the pure neural network system. Also, the
optimal selection of the most influencing variables was performed successfully by the
hybrid intelligent system. The proposed artificial intelligent systems could be adopted
on-line as a reliable controller of the thermal power plant boiler.