Hybrid onboard model of high-flow dual variable cycle engine based on deep reinforcement learning

Real-time onboard performance modeling for aero-engines holds a pivotal position in engine control and fault diagnosis. Nevertheless, the relentless pursuit of enhancing aero-engine performance has led to increasingly intricate engine architectures, posing significant challenges to the real-time execution of onboard models. To improve the real-time performance while maintaining accuracy, and address the inherent problem that traditional models can only start to simulate at a fixed point, an aero-engine hybrid onboard model based on deep reinforcement learning is proposed. This study focuses on the high-flow dual variable cycle engine, firstly delving into the engine’s thermodynamic principles comprehensively, then establishing and validating the Component Level Model (CLM), which serve as the inner loop of hybrid model. Subsequently, Twin Delayed Deep Deterministic (TD3) policy gradient algorithm is integrated to establish the outer loop of hybrid model within full flight envelope and operating states. Simulation results show that the hybrid model has the characteristics of faster convergence, higher accuracy and superior real-time performance. Finally, the inner-outer loop design enables hybrid onboard model to approach the domain of solution quickly, significantly reducing the calculation time of simulation initialization and ensuring model convergence throughout transition processes.