Fault Detection Relevant, Neural Network and Evolutionary Algorithm based Model for a Single-shaft Industrial Gas Turbine

Analytical redundancy based fault detection system generally requires an accurate model. Meeting this requirement for a system like an industrial gas turbine is quite difficult as it involves nonlinear energy and momentum exchange between the working fluid and stationary and rotating parts of the system. In the latest designs the gas turbine is equipped with adjustable guide vanes (IGV) and variable stator vanes (VSV) –located in the compressor section of the system –to effect in better performance at part load and to improve stall characteristics during starting. This adds up to more nonlinearity of the system. In this paper the result of an attempt to develop a substitute nonlinear model based on multilayer neural network (MLNN) and evolutionary algorithm (EA) for a single-shaft gas turbine having IGVs and VSVs is presented. Included are calculation of MLNN topology and parameters and calculation of model confidence intervals (CI) based on two assumptions –whole weight and bias parameters, and last layer parameters. Real operation data collected in 10 seconds interval are used for training as well as validation of the model. A test on the capability of the model for fault detection revealed that it is possible to detect inlet temperature changes as low as 5 deg. C. The model can also be used for performance optimization studies.