Comparative Study on Predictive Tool for Cyclone Separation Efficiency using Machine Learning Algorithm

Cyclone separation efficiency is one of critical factors analysed in designing cyclone separators; in which its efficiency is dependent on design parameters such as cyclone separator dimension, operating temperature, and particle density. This report aims at investigating the complex stochastic relationship between the separation efficiency of particle density and geometrical dimension (tuning ratio) while maintaining similar size of cyclone separator and operating temperature. Experimental data shows that the separation efficiency behaves unstably at particle size of less than 0.3 μm and demonstrates stable behavior at particle size beyond 0.3
${\upmu}\mathbf{m}$
, and thus, the feasibility of using Machine Learning Algorithm (MLA) is explored. Comparative studies on various nonlinear regression models, e.g., Neural Network (NN), Boosted Regression Tree (BRT), etc. are presented. Optimal parameters for each model are tuned using manual search to find the best suited method. It is found that BRT has the lowest error accuracy with an average MAE of 0.349 while NN shows the most reliable performance with a standard deviation of 0.1557.