Hybrid acceleration schedule design for gas turbine engine using adaptive sample error weighting multilayer perceptron network

Acceleration schedules are pivotal for gas turbine engines by providing appropriate control references to enhance acceleration performance. Traditional corrected parameter based (CPB) methods are prompt but cannot exploit engine acceleration performance across the flight envelope, while data-driven methods enhance acceleration performance but fall short in real-time performance. This paper proposes a hybrid acceleration schedule (HAS) integrating the CPB and data-driven compensation module to balance real-time and acceleration performance. Firstly, the CPB establishes a basic low-precision acceleration schedule. Then, the multilayer perceptron network with adaptive sample error weighting (ASEW) strategy is trained on CPB residual data as compensation module. The hybrid framework allows for precise reconstruction of CPB residual data via small-sized multilayer perceptron network, preventing severe degradation in real-time performance. The ASEW enhances compensation precision by dynamically selecting hard-to-train samples and adjusting their weights. The HAS achieves acceleration schedule mean error of 0.89 %, comparable to data-driven methods (about 1 %) and outperforming traditional method (2.55 %), guaranteeing satisfactory acceleration performance across flight envelope. It also ensures real-time performance comparable to CPB and reduces prediction time from about 12.7 ms in data-driven method to 1.66 ms. Moreover, the ASEW strategy effectively reduces the maximum prediction error while keeping the mean error nearly unchanged.