In combined cycle power generation using gas turbines, two cycle parameters viz. CDP and TIT are very important. The loss of CDP measurement leads to loss of few megawatts of generation and the non-availability of TIT, which cannot be measured; puts great limitations on efficient turbine operation and maintenance planning. The aim of this work is therefore to soft sense these cycle parameters of the gas turbine using Artificial Intelligence for online use in a power plant.

There are many ways in soft computing to develop an input/output mapping for such applications, with some methods offering advantages over others. Three different types of soft computing techniques have been used in this work viz., a feed forward NN model using BP algorithm, a type-1 FLS and a type-2 FLS were used in both the problems and their performance compared based on ability to handle uncertainties.

A soft sensor for CDP is developed using five measurable plant parameters having influence on CDP and using them as an input vector for the design of NN and FLSs. The NN / FLS once trained can be used as a soft sensor for online estimation of CDP, which can be used in the control loop in case of loss of CDP measurement.

The data collected for training and testing of the NN / FLS designed consist of plant parameters selected as inputs covering different operating conditions of the turbine and seasonal changes in atmospheric conditions throughout the year.

Since TIT cannot be directly measured so it is first estimated offline by applying heat balance through the combustion chamber using a numerical model. For validation of target value of TIT, a model of the turbine using infinitesimal stage efficiency method is developed. Proceeding forward starting from the TIT value obtained and arriving at the exhaust condition; the validation of computed TIT is done through exhaust temperature measurement. The TIT thus computed serves as the target value for NN and the FLSs. A soft sensor is then developed for TIT using NN, type-1 and type-2 FLS.

This thesis presents a comparison of the three designs applied to both the CDP and the TIT problem and shows that a type-2 fuzzy logic system is more robust in the presence of uncertainties than NN or type-1 FLS. This is because the uncertainties are being incorporated in the rules of the FLS when type-2 fuzzy sets are used. Though mathematically more complex and difficult, a type-2 FLS provides additional degree of freedom to model uncertainty.