Mean line aerodynamic design of an axial compressor using a novel design approach based on reinforcement learning

Abstract

This paper develops a novel design approach based on reinforcement learning, which can independently complete the mean line aerodynamic design process of the axial compressor. The approach combines Deep Deterministic Policy Gradient (DDPG) algorithm with mean line aerodynamic predicting program HARIKA to acquire the design experiences of the axial compressor. DDPG combines basic reinforcement learning algorithm with artificial neural networks to get continuous observation and give corresponding actions. After the specific modification of the DDPG, multi-objective optimization can be integrated into the design process. Under the guidance of this approach, the design and optimization processes of a 9-stage high-pressure axial compressor were completed without expert experiences. At the design point, the isentropic efficiency was 88.5% and the surge margin was 25%, which meets the requirement of the compressor’s efficiency and stability. And there was an increase of 13.4% and 22%, respectively, compared to the initial design. Moreover, through the analysis of the design results, the distributions of aerodynamic parameters conform to expert experiences. To verify the approach, traditional optimization methods, multi-island genetic optimization algorithm (GA), and multi-objective particle swarm optimization algorithm (MOPSO) were used to solve the same optimization problem. The DDPG optimized efficiency was 0.2% lower than the traditional optimization method, and the pressure ratio at the work point was, respectively, 0.6% and 2% higher than that of the MOPSO and GA, which proved the effectiveness of the new design approach. Furthermore, after training, this approach can give design results immediately near the specific design requirements, which is different from the traditional optimization methods. The new approach saved 93% evaluation steps compared to the GA in the −3% design point and finished the design process in 8 steps in the +3% design point, where GA failed to complete.