Document
Transient Optimization of a Gas
Turbine Engine
2023 AIAA SciTech Forum Jonathan Kratz NASA Glenn Research Center th January 24 , 2023 1/24/2023 1 2023 AIAA SciTech Forum
Objectives
• Apply optimization techniques to optimize fuel flow control transient *
limit logic for a turbofan engine to maximize operability to enable
better performance
• Evaluate the optimized solutions against a baseline
• Leverage the optimization results and sensor feedback to guide a
machine learning algorithm to modify the transient limit logic in order
to achieve the best performance on the real system
* Transient refers to a change in engine power/thrust demand associated with acceleration or deceleration of the engine shafts.
1/24/2023 2023 AIAA SciTech Forum 2
Contents and Organization
• Background
• Description of the Approach
• Transient limit logic optimization • Optimization over the lifespan of the engine
• Application for Illustrating the Approach
• Transient Optimization Results (Abbreviated For Video Presentation)
• Engine Lifespan Optimization Results
• Conclusions
1/24/2023 2023 AIAA SciTech Forum 3
Background
• Challenge – maintaining operability Altitude during engine power transients throughout a vast operating envelope Mach Number • Transient limit logic • protects the engine from compressor stall/surge and combustor blowout • Transient limit logic accounts for nearly 75% of the total time dedicated to control † system development • Is often implemented as a shaft acceleration/deceleration schedule or a ratio unit (fuel flow / compressor discharge pressure) schedule • Control design is guided by requirements for thrust responsiveness and adequate operability margin • Engine design and associated performance are constrained by operability requirements † *Image Credit to NASA and the Tool for Turbine Engine Reference: Jaw & Mattingly, Aircraft Engine Controls: Closed - loop Transient Analysis ( TTECTrA ) Design, System Analysis, and Health Monitoring, 1/24/2023 2023 AIAA SciTech Forum 4
Background (cont.)
REAL ASSET • Common techniques for transient limit logic design are sub - optimal • Engine performance shifts with degradation and maintenance → an optimal design is only optimal for a specified health state • Digital twin technology can be leveraged to design and potentially update controls Image Credit: NASA
=
• Optimization techniques could be applied to refine the transient limit
1 0 0 1 0 1 0 1
logic
0 1 1 1 0 0 1 0
• Machine learning can be applied to update the logic as the engine ages
1 0 0 1 0 1 1 1
to maintain near optimal dynamic performance DIGITAL REPRESENTATION 1/24/2023 2023 AIAA SciTech Forum 5
The Optimization Approach
• A genetic algorithm is applied to identify the “optimal” fuel flow rate profile that maximizes operability as defined by the transient stack usage (TSU) • applies functions of elitism, carry - over (replication), cross - over (reproduction), and immigration • utilizes rank - based selection with probabilities based on a pareto distribution • The inputs are the fuel flow va lues at various 𝑃𝑅 − 𝑃𝑅 𝑆𝑆 𝑇𝑆𝑈 = 𝑚𝑎𝑥 × 100% times throughout the transient, constrained to 𝑃𝑅 − 𝑃𝑅 𝑠𝑡𝑎𝑙𝑙 𝑆𝑆 be monotonically increasing or decreasing SS = Steady - State • An iterative root solving technique is leveraged to stretch/compress the fuel flow input profile PR = pressure to achieve the desired thrust response time ratio • Use optimized results to derive transient limit W = corrected c schedules w Data points f flow rate defining w f w = fuel flow f profile rate Time 1/24/2023 2023 AIAA SciTech Forum 6
The Engine Lifespan Optimization Approach
• Use the digital twin to • Design a conservative schedule for an end - of - life (EOL) engine • Design an aggressive schedule for a new (NEW) engine • Create an array of discrete schedules between the two extremes • Could update the “aggressive schedule” and discrete options over the lifespan of the engine • Use a reinforcement learning (RL) algorithm to shift the schedule in small increments and accumulate rewards based on sensor feedback 1/24/2023 2023 AIAA SciTech Forum 7
The Engine Lifespan Optimization Approach
• Start at the conservative schedule and march toward the aggressive schedule • Objective: minimize thrust response time while respecting compressor operability margin constraints • Uses a Q - Learning algorithm • Positive rewards for: • Reducing thrust response time • Increasing operability if the operability limit was violated with the prior action • Negative rewards for: • Violating the operability constraint MID = Mid - life 3Q = 3 - Quarter life • Increasing the thrust response time while not violating the operability limit Do Not Exceed Line was defined • Staying on the same schedule based on the EOL running line 1/24/2023 2023 AIAA SciTech Forum 8
Application - AGTF30 Engine
• Conceptual two - spool geared turbofan
Advanced Geared Turbofan
• Produces ~30,000 lb of thrust at sea f
30,000 lb (AGTF30)
level static (SLS) conditions f • Envisioned for single - aisle applications • Included advanced technologies • Compact core • Variable area fan nozzle • MATLAB/Simulink® model developed with the Toolbox for Modeling & Analysis of Thermodynamic Systems (T - MATS) • Includes a baseline controller with representative performance • Includes engine health parameters 1/24/2023 2023 AIAA SciTech Forum 9
Optimization Results - Acceleration
Reduced operability stack usage by 31% 1/24/2023 2023 AIAA SciTech Forum 10
Engine Lifespan Optimization Results
New Engine Training
Reduced thrust response time by nearly 0.5s or 11.8%
1/24/2023 2023 AIAA SciTech Forum 11
Engine Lifespan Optimization Results
Degraded the engine from new to
• RL adjusts as the engine degrades
mid - life to see how the RL agent
• Attempted to refine the schedules
would adjust
by updating the “optimized”
aggressive schedule as the engine
ages.
• Only resulted in a 0.02s improvement • Updating the schedule options may not add enough benefit to be warranted 1/24/2023 2023 AIAA SciTech Forum 12
Conclusions
• A genetic algorithm has been applied to optimize transient limit logic for a turbofan engine • Reduced the use of the overall operability stack by 31% compared to a baseline controller • Demonstrated success with generalizing optimization results across a wide range of flight conditions* • Demonstrated how the approach can be modified to optimize for different goals (ex. reduced peak temperatures to preserve engine life)* • A reinforcement learning (RL) algorithm was applied in an approach to adjust the transient limit logic of a turbofan engine over its lifespan • Demonstrated the ability reduce the thrust response time by 11.8% while respecting operability limits • Results suggest the need to update the “optimal” schedule as determined by the optimization approach in concert with a digital twin might be unnecessary, thus reducing the workload for implementation • Potential directions for future work • Address challenges for practical implementation of the RL approach • Modify the goals of RL approach to be better aligned with commercial engine applications * Results omitted to meet video time constraint 1/24/2023 2023 AIAA SciTech Forum 13
Acknowledgments
• Funded by the Transformational Tools & Technologies (TTT) project
under the Aeronautics Research Mission Directorate (ARMD)
Questions/Discussion
Contact Information
Jonathan Kratz – jonathan.kratz@nasa.gov
1/24/2023 2023 AIAA SciTech Forum 14