MC-UVE-GA-PLS算法用于精馏软测量辅助变量选择

Secondary variable selection in distillation column soft sensor using MC-UVE-GA-PLS algorithm

现代工业生产的复杂化使得候选辅助变量增多,大量的输入变量会使软测量模型过拟合,影响模型的预测效果。针对这个问题,提出了一种蒙特卡洛无信息变量消除结合遗传算法偏最小二乘(MC-UVE-GA-PLS)选择辅助变量的方法。该方法在运用GA算法搜索最优变量子集之前,采用MC-UVE方法消除与模型不相关的变量,使GA算法能有效地搜索出对响应变量预测贡献最大的变量子集。用本文提出的方法建立了工业精馏塔浓度软测量模型,仿真结果表明本文提出的辅助变量选择方法不仅能提高模型的预测能力,而且能简化模型的复杂性。

The complication of modem industrial production leads to the increase of the number of candidate secondary variables. A large number of input variables may overfit the soft-sensor model and impact prediction of model. To solve this problem, an algorithm based on Monte Carlo Uninformative Variable Elimination Genetic Algorithm Partial Least Squares （MC-UVE-GA-PLS） is proposed to select secondary variables. In the proposed approach, MC-UVE method is adopted to eliminate the variables with irrelative information for modeling before applying GA to search optimal variables. It is effective for GA to select variables which has maximum prediction contribution to response variable. The proposed method has been applied to develop a soft-sensor model for the concentration of industrial distillation column and the result demonstrates that the method can not only improve the prediction ability of model but also simplify the model complexity.