Multi-scale linear programming support vector regression for ethylene distillation modeling

In the ethylene distillation process, ethylene concentration fails to be effectively controlled if lacking theoretical guidance. To a certain extent, the neural network method can estimate and control ethylene concentration, but there are some limitations, such as overfitting and low reliability. In this paper, a hybrid algorithm is proposed for the soft sensing modeling of ethylene distillation column based on the v-multi-scale linear programming support vector regression and particle swarm optimization. In the hybrid algorithm, estimation function is composed of a linear combination of a series of feature spaces, which is optimized by linear programming, and particle swarm optimization is used effectively for the regression parameters selection. Numerical simulations further demonstrate that the algorithm has great effectiveness in the modeling for ethylene distillation.