Research of Cyclone Optimization Based on CFD, GMDH-Type Neural Network and Genetic Algorithm

Abstract

Gas cyclone has two main parameters for evaluating separation performance, separation efficiency and pressure drop through cyclone. They are closely influenced by the geometrical design variables of the cyclone. This study performed optimization of the cyclone performance for the cyclone shape based on computational fluid dynamics (CFD), GMDH type neural network and genetic algorithm (GA). First, CFD was used to obtain the data of the cyclone performance parameters. As result of the CFD validation, the errors of the reference model and CFD were 0.5 % and 2 % for the pressure drop and the separation efficiency. Secondly, the meta-model of the cyclone performance was derived by using GMDH algorithm based on supervised learning of machine learning. The fitness of the modelling results was shown using the correlation coefficient. As results of the GMDH, the correlation coefficients of meta models of the separation efficiency and the pressure drop were 98.9 %, 99.7 %, respectively. Finally, we performed optimization of the meta model by applying GA. When the optimal point was compared with the reference model, the performance of the optimal point was improved by 24.31 % and 8.32 % for pressure drop and the separation efficiency, respectively.