Optimization of Combustor Based on Machine Learning Adaptive Dimensionality Reduction Method

Gas turbine is an important power equipment in modern industry, and the combustor is its key component. In order to achieve high efficiency combustion, complex geometric structures are sophisticatedly designed. The swirler has a significant impact on its performance. Excellent geometric design can improve combustion efficiency, reduce pollutant emissions, and maintain outlet temperature uniform, thereby extending the service life of turbine components. This research studies the typical geometric structures of combustor swirlers and uses machine learning method to improve performance. The study employed a real combustor as a baseline. Rigorous CFD methods, validated by experimental data, were adopted. The following analysis are based on the similar computational settings. To achieve multi-parameter optimization of geometry, an automated platform was constructed, making large number of parametric computations accessible. In this paper, totally 10 parameters were selected as initial inputs, and 170 sets of data were generated through Latin Hypercube Sampling. Subsequently, several output parameters, including OTDF (Overall Temperature Distribution Factor) and RTDF (Radial Temperature Distribution Factor), were compared. Considering the Pearson coefficient and combustion efficiency comprehensively, RTDF was chosen as the optimization objective. Based on the data obtained from calculations, a high-precision DNN (Deep Neural Network) was established to map relationship from inputs to outputs. To identify key sensitive parameters, an adaptive dimensionality reduction method was developed. By considering SHAP (SHapley Additive exPlanation) value, an iterative process of elimination and searching were conducted. As a result, the high-dimensional inputs were reduced to only the most important six feature parameters, which is 40% reduction of inputs. Finally, optimization was performed using a genetic algorithm, which reduced the RTDF from 1.21 to 0.36, marking a 70% decrease. Additionally, a flow field analysis of the optimal results was conducted, and the increase in vorticity also revealed the flow mixing physics.