The prediction of ethylene cracking furnace yields on line is significant in industrial production for advanced control and energy efficiency. Due to the great differences between different operating conditions

single condition and modeling may not satisfy the requirement of practical process. Considering the similarity of cracking furnace and the reduction of acquisition cost

the history data are utilized assisting transfer learning to improve the accuracy of operating condition classification. Least squares support vector machines (LS-SVM) is employed in modeling cracking furnace yields in different operating conditions

which enjoy stronger generalization ability and faster convergence speed compared with standard SVM. The accuracy is further improved by optimizing parameters of LS-SVM with particle swarm optimization (PSO)

and thus establishing different operating condition models for yields prediction. The simulations and operating condition classifications are given based on the real industrial data to demonstrate that the operating condition classification is more reasonable. The prediction of LS-SVM optimized with PSO is more accuracy and behaves good trend tracking performance.