A general methodology has been established to set up GT based plant simulators to perform analysis and to identify inverse model parameters. The attention is focused on three categories of inverse problems faced in setting up the plant simulator: i) sizing of components; ii) calibration on the basis of test acceptance data; iii) actual status recognition from data collected by the plant monitoring system. Due to the different nature and requirements of the above problems different solution approaches have been adopted: hybrid stochastic-deterministic algorithms for model calibration and neural techniques for status recognition. The methodology to set up and solve models of GT based CHP plants has been illustrated and discussed. The plant model is characterised by the introduction of two kinds of functions: a) Reality Functions, which calibrate the model to reproduce accurately the N&C plant behavior; b) Actuality Functions, which update the map of each component to predict actual deteriorated plant operations. The identification of RFs and AFs coefficients poses two inverse engineering problems which ask for different strategies and solution techniques. In general, RF identification from available acceptance test data leads to a problem of error function minimisation. Among the applicable solution techniques those based on hybrid Evolutionary-Deterministic Algorithms have been selected. The application of a hybrid technique GA-ECRQP brought to good results in terms of accuracy in reproducing the plant N&C behavior. Component status recognition (i.e. identification of AF coefficients) is based on a single measured point because it has to be done in real time using DCS data. In order to meet the requirement of a fast and sufficiently accurate solution Neural Network (NN) baed techniques have been proposed and successfully applied. NN status recognition performance has been compared to that achieved by the inverse solution of the plant physical-empirical simulator. Errors introduced by NN in AF estimation are smaller than 1%. This satisfactory accuracy and the really short computational time required (some 40 μs) show the potentialities of the NN approach for on-line or quasi on-line applications to support plant management decisions.
Cerri, G., Borghetti, S., Salvini, C. (2005). Inverse Methodologies for Actual Status Recognition of Gas Turbine Components. In ASME Conference Proceedings POWER2005 (2005) (pp.299-306). New York, N.Y. : Mechanical Engineering magazine/ASME.
Inverse Methodologies for Actual Status Recognition of Gas Turbine Components
CERRI, Giovanni;SALVINI C.
2005-01-01
Abstract
A general methodology has been established to set up GT based plant simulators to perform analysis and to identify inverse model parameters. The attention is focused on three categories of inverse problems faced in setting up the plant simulator: i) sizing of components; ii) calibration on the basis of test acceptance data; iii) actual status recognition from data collected by the plant monitoring system. Due to the different nature and requirements of the above problems different solution approaches have been adopted: hybrid stochastic-deterministic algorithms for model calibration and neural techniques for status recognition. The methodology to set up and solve models of GT based CHP plants has been illustrated and discussed. The plant model is characterised by the introduction of two kinds of functions: a) Reality Functions, which calibrate the model to reproduce accurately the N&C plant behavior; b) Actuality Functions, which update the map of each component to predict actual deteriorated plant operations. The identification of RFs and AFs coefficients poses two inverse engineering problems which ask for different strategies and solution techniques. In general, RF identification from available acceptance test data leads to a problem of error function minimisation. Among the applicable solution techniques those based on hybrid Evolutionary-Deterministic Algorithms have been selected. The application of a hybrid technique GA-ECRQP brought to good results in terms of accuracy in reproducing the plant N&C behavior. Component status recognition (i.e. identification of AF coefficients) is based on a single measured point because it has to be done in real time using DCS data. In order to meet the requirement of a fast and sufficiently accurate solution Neural Network (NN) baed techniques have been proposed and successfully applied. NN status recognition performance has been compared to that achieved by the inverse solution of the plant physical-empirical simulator. Errors introduced by NN in AF estimation are smaller than 1%. This satisfactory accuracy and the really short computational time required (some 40 μs) show the potentialities of the NN approach for on-line or quasi on-line applications to support plant management decisions.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
