Employment of Statistical and Artificial Intelligence Techniques for Prediction of Combustion Dynamics in an Experimental Swirl-stabilized Combustor

In the present work, the relations between three output quantities measured in an experimental combustor and two input quantities of overall equivalence ratio (φ) and secondary fuel injection rate (Qsec) were studied. The three measured output quantities of the combustor are the amount of NOx emission, noise level, and the level of pressure oscillations generated by the combustion field. The experimental combustor has certain applications in power plant gas turbines. In order to study the relations between the measured output and input quantities of the combustor, two different data mining approaches were utilized. Specifically, in this research, Multilayer Perceptron (MLP) Neural Network and Response Surface Method (RSM) techniques were employed to provide an estimation of the nonlinear relationship between the inputs and outputs of the combustor. The related experiments were already performed using four different types of secondary fuel injectors (with different designs) for an overall equivalence ratio between φ =0.7~0.9, along with different amounts of secondary fuel injection rate in the range of Qsec=0.6~4.2 l/min. The results show that, in general, both the MLP neural network and RSM approaches have good predicting capability for estimation of noise level, level of pressure oscillations, and NOx emission. However, the degree of agreement between the predicted and measured values would even be enhanced for the case of NOx emission. Also, comparing the two data mining methods, the results indicate that the MLP neural network has better prediction ability for estimation of various combustor parameters than the RSM, for the cases of different injectors.