Wide-Range Variable Cycle Engine Control Based on Deep Reinforcement Learning

In this paper, a controller design method based on deep reinforcement learning is proposed for a wide-range variable cycle engine with a turbine interstage mixed architecture. The PID controller is subject to limitations, including single-input single-output limitations, low regulation efficiency, a...

Full description

Saved in:
Bibliographic Details
Published in:Aerospace Vol. 12; no. 5; p. 424
Main Authors: Ding, Yaoyao, Wang, Fengming, Mu, Yuanwei, Sun, Hongfei
Format: Journal Article
Language:English
Published: Basel MDPI AG 01.05.2025
Subjects:
action space pruning
Aircraft
Algorithms
Artificial intelligence
Control systems design
Controllers
Deep learning
deep reinforcement learning
Design
Efficiency
Energy consumption
Engine control
Engines
Flight conditions
Geometry
Laws, regulations and rules
Methods
Neural networks
PID control
Process controls
Proportional integral derivative
SISO (control systems)
Turbines
variable cycle engine
Variable cycle engines
Variables
wide-range
Massachusetts
ISSN:2226-4310, 2226-4310
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In this paper, a controller design method based on deep reinforcement learning is proposed for a wide-range variable cycle engine with a turbine interstage mixed architecture. The PID controller is subject to limitations, including single-input single-output limitations, low regulation efficiency, and poor adaptability when confronted with contemporary variable cycle engines that exhibit complex and multi-variable operating conditions. To solve this problem, this paper adopts a deep reinforcement learning method based on a deep deterministic policy gradient algorithm, and it applies an action space pruning technique to optimize the controller, which significantly improves the convergence speed of network training. In order to verify the control performance, two typical flight conditions are selected for simulation experiments as follows: in the first scenario, H = 0 km and Ma = 0, while in the second scenario, H = 10 km and Ma = 0.9. A comparison of the simulation results shows that the proposed deep reinforcement learning controller effectively addresses the engine’s multi-variable coupling control problem. In addition, it reduces response time by an average of 44.5%, while maintaining a similar overshoot level to that of the PID controller.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:2226-4310
2226-4310
DOI:10.3390/aerospace12050424