Least Squares-Support Vector Regression for Determining Product Concentrations in Acid-Catalyzed Propylene Oligomerization
2018
Sivaramakrishnan, Kaushik | Nie, Jinjun | Klerk, Arno de | Prasad, Vinay
This work is concerned with the development of multivariate calibration models to establish spectrum-composition relationships for the hydrocarbon products in the H-ZSM-5-catalyzed oligomerization of propylene. Regression models based on two multivariate methods were investigated in this work: least-squares-support vector machines (LS-SVM) and partial least-squares (PLS) regression. The performance of two nonlinear kernels, radial basis function (RBF) and polynomial, is compared with PLS as well as its variant, interval-PLS regression (i-PLSR). For comparison with i-PLSR, the Fourier transform infrared (FTIR) spectra of the products served as inputs and the respective C₁–C₁₀ concentrations, obtained from gas chromatography (GC), were the outputs. The sensitivity of the product distribution to inlet operating conditions was also evaluated through the calibration methods. Spectral clusters having distinct chemical character were identified using principal component analysis (PCA) and hierarchical clustering analysis (HCA) and also used as inputs to the different regression techniques to compare with the full spectrum models. It was found that the best performing spectral regions from i-PLSR had chemical relevance and agreed with findings from HCA, improving the predictive capabilities significantly. The decreasing order of performance of the chemometric methods evaluated was: LS-SVM-RBF > LS-SVM-Polynomial > i-PLS > PLS. The prediction accuracy of RBF kernel-based LS-SVM regression technique was the highest, indicating its suitability for effective online monitoring of moderately complex processes like acid-catalyzed propylene oligomerization.
显示更多 [+] 显示较少 [-]