Prediction and sensitivity analysis of pressure pulsation in a LDI combustor based on a fully connected neural network

This paper presents a pressure pulsation prediction model and conducts sensitivity analysis on pressure pulsation. Experiments are performed on a multi-swirl lean direct injection (LDI) combustor under varied operating conditions to establish a comprehensive dataset. The raw data undergo fast Fourier transform (FFT) and phase space reconstruction to determine combustion mode (stable/unstable), dominant frequency, and amplitude. A fully connected neural network (FCNN) is utilized to predict these pressure parameters, and it demonstrates superior performance compared to support vector regression (SVR) and random forest (RF) models. For test data, the FCNN achieves an average relative error (MRE) of 2.54 % for dominant frequency and a mean absolute error (MAE) of 0.64 % for pressure amplitude, suggesting high accuracy and generalization ability. Furthermore, the FCNN model is assessed for its alignment with physical laws by varying input features. Results indicate that the FCNN model exhibits the best physical consistency among the models. Employing the FCNN model as the surrogate, Sobol’ sensitivity analysis identifies the fuel-air ratio as the most influential parameter, with significant impacts also from inlet flow rate and inlet temperature, while nozzle position exerts a minor influence. Additionally, individual parameter effects on combustion instability are minimal, and instability is primarily driven by parameter interactions.