Performance diagnoses of heavy-duty industrial gas turbines often rely on measured data from on-site monitoring systems (OSM), subjected to larger uncertainties and possible biases. The measured data are used to analyze gas turbine heat balance and estimate immeasurable performance characteristics such as firing temperature and component health parameters. Traditional heat balance techniques are deterministic, and, thus, calibration uncertainty is not mitigated. In this paper, a method of model-based data reconciliation (MBDR) and bias detection was developed, serving as a probabilistic process of reducing calibration uncertainty while eliminating contamination effects caused by measurement biases. This method utilizes physics-based gas turbine models to reconcile multiple data sets while the model health parameters are inferred simultaneously. Levenberg–Marquardt algorithm was utilized to solve the maximum-likelihood problem, i.e., minimizing Least Squares. A hypothesis test scheme using sequential bias compensation was utilized for bias detection and neutralizing smearing effects. To reduce the computation time in MBDR and bias detection, the Response Surface Methodology (RSM) was applied to generate surrogate model. A systematic way of data selection using Multiscale Principal Component Analysis was also employed, serving as an efficient way of filtering large data sets for the use of MBDR. This proposed methodology was demonstrated by application to GE 7FA gas turbines. Results showed significant reduction in calibration uncertainty and smearing effects.
Semi-nonparametric approach for measured data reconciliation based on the Gram-Charlier series expansion
