Deep reinforcement learning method for turbofan engine acceleration optimization problem within full flight envelope

Abstract

In order to solve the multidimensional restricted optimization problem of aeroengine acceleration process and improve the acceleration performance of full envelope, a design method of aeroengine full envelope acceleration controller based on deep reinforcement learning is proposed. Firstly, the thermo-aerodynamic principle of the engine acceleration process is analyzed, and the mathematical analytical model of the multidimensional restricted optimization problem of the turbofan engine acceleration process is established. On this basis, the design method of acceleration controller based on twin delayed deep deterministic policy gradient is studied. The proportional integral controller is designed to pursue the asymptotic stability at the acceleration destination and stable control is achieved through switching. Additionally, the flight envelope is divided into regions by clustering to reduce the span of data changes, so as to enhance the network convergence ability of deep reinforcement learning. Finally, based on the similarity conversion method of equal engine inlet total temperature, the acceleration controller of full flight envelope is realized. The digital simulation results demonstrate that compared with the traditional proportional integral differential controller, the acceleration controller based on twin delayed deep deterministic policy gradient shortens the acceleration time up to 48.33%, and the control effect in a large flight envelope (H=0∼10 km, Ma=0∼1.6) has been verified.