Using Gaussian Processes to model combustion dynamics

Modelling the dynamics of combustion is a challenging task due to the non-linear interaction of many processes involved, including chemical kinetics, flame dynamics and acoustic pressure variations inside the chamber. Given that gas turbine engines are the dominant power generation sources, more sophisticated models that can make accurate and reliable predictions regarding the combustion processes and its efficiency, are always in high demand. This paper discusses the development of a data-driven model that is based purely on experimental data, collected from a combustion test rig. The model has been developed using Gaussian Processes, an advanced Bayesian non-parametric machine learning algorithm. The collected data, including pressure inside the combustion primary zone and structural vibration, were all considered as possible candidates for adapting this algorithm to the dynamical characteristics of the combustion chamber under investigation. Accuracy in prediction using this empirical model was investigated for different combinations of experimental data and fractions of them, using the root mean squared error as performance measure. The covariance function parameters of the Gaussian Process model were optimised using a gradient-based algorithm for the best possible adaptation to the experimental dataset.