Date of Award
8-16-2024
Document Type
Open Access Dissertation
Department
Mechanical Engineering
First Advisor
Jamil Khan
Abstract
The growth in renewable energy sources and retirement of large baseload coal-fired power stations has led to an accompanying decrease in reliability and security of the electrical grid. Since renewable energy sources are typically non-dispatchable, this can lead to blackouts and/or brownouts for customers. Heavy duty gas turbine power plants (HDGT) offer a solution to this problem. HDGTs are dispatchable, clean, and offer flexibility in the fuel they consume, but operational limitations must be well understood to fully exploit their benefits.
One of the main operational limitations is the tip clearances in the gas turbine. In many cases, the gas turbine operability is limited based on worst-case scenarios. These operability restrictions can be avoided in most circumstances provided the tip gaps are known in real time. Sensors have been developed to measure tip gaps, but are extremely expensive, imprecise, difficult to install, and unreliable.
This work presents a collection of reduced-order models that are built on supervised and unsupervised machine learning concepts. First, the work provides a review of unsupervised clustering techniques and highlights the strengths and limitations of each for the determination of operational clusters. The work then provides a means to estimate the operational cluster from limited measurements and predict unmeasured quantities. Using these model outputs, estimates for cooling/heating effectiveness and thermal time constants are obtained. The results are combined to provide an open-loop thermal-mechanical state for each component of interest and its corresponding deflection. The deflections are properly combined to determine the transient real-time clearances. Finally, with the use of the Unscented Kalman filter and a limited number of temperature measurements, the model is refined to provide the most accurate estimate. The model results are compared to full-scale HDGT test measurements and are found to be extremely accurate. Additionally, the models are more reliable than the clearance probe sensors.
Rights
© 2024, Donald Earl Floyd
Recommended Citation
Floyd, D. E.(2024). Application of Machine Learning Techniques and the Unscented Kalman Filter to Real-time Gas Turbine Clearance Prediction. (Doctoral dissertation). Retrieved from https://scholarcommons.sc.edu/etd/7726