A Methodology for Sampling with a Pattern Classifier in Gas Turbine State Space to Create Transient Surrogate Models
Loading...
Author(s)
Ozcan, Metin Firat
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
Gas turbine technologies are critical for reducing fuel consumption, emissions, and noise. A suite of
technologies is necessary to reach the environmental impact goals. Therefore, accurately selecting the
technology portfolio is crucial. Conventionally, gas turbine technology evaluations depend on steady-
state performance because steady-state performance can accurately approximate each phase in an
aircraft mission profile and yield an aircraft size estimate at the end of the mission analysis. Since
aircraft size is the primary effect on aircraft performance, this approach works well to evaluate many
technologies. However, some technologies, such as variable area fan nozzles and highly loaded
compressors, require safe transitions in operation, but a steady-state model cannot model performance
changes over time. Therefore, transient gas turbine models are necessary for evaluating such
technologies thoroughly.
Among several choices in the literature for transient gas turbine models, data-driven alternatives, such
as Piecewise Linear Models (PLMs) and transient surrogate models, have appealing properties, like
computational speed and relating gas turbine performance to only measurable variables, for evaluating
technologies with transient behavior. Moreover, transient surrogate models provide analytically
differentiable relationships between the performance metrics and measurable variables. Therefore, this
thesis focused on transient surrogate models.
Transient surrogate model accuracy depends on its sampling. The conventional approach of
independently varying sampling variables can cause up to ninety-eight percent of failed cases with a
transient gas turbine model. Therefore, this thesis proposed using a pattern classifier during the
sampling process to filter the infeasible points before spending computational time executing them.
Two large-scale samplings at sea level static with the developed direct-drive and geared fan gas turbine
models showed that pattern classifiers filtered more than sixty-five percent of the infeasible points
while correctly classifying more than ninety-seven percent of the feasible points.
Transient surrogate models approximated the two large-scale transient samplings. Except for the low-
pressure compressor's stall margin, the surrogate models matched or were more accurate than the
PLMs, when predicting the performance metrics of interest, like net thrust, burner exit total
temperature, and high-pressure compressor stall margin. Neither the PLMs nor the surrogate models
accurately predicted the low-pressure compressor's stall margin during aggressive maneuvers.
Sponsor
Date
2023-04-30
Extent
Resource Type
Text
Resource Subtype
Dissertation