基于即时学习与输出相关的变量加权研究

Research on Variable Weighting of Output-Related Based on Just-in-Time Learning

即时学习算法是根据某种最优准则,从历史数据中选出容易检测并且与主导变量密切相关的辅助变量,从而实现对主导变量的预测。局部加权偏最小二乘算法(LW-PLS)是即时学习算法中常用的建模方法,而相似性样本的选择是LW-PLS能否取得良好建模效果的最关键因素。众多试验研究已经证明,基于回归系数和相关系数给变量加权的即时学习算法能明显提高模型的预测精度,而变量权重的确定也会影响预测效果。因此,在基于与输出相关的即时学习算法的基础上,探讨了不同的权重函数和权重系数的不同阶次对建模效果的影响,并分别在数值例子和硫回收单元实际例子中进行了验证。结果证明,在实际工业过程中,合理地选择权重函数,并且选择合适的权重次数,能够明显提高系统的预测性能。

The just-itime learning algorithm selects auxiliary variables from historical data that are easy to detect and closely related to the dominant variables according to some optimal criteria,so as to realize the prediction of the dominant variables.Locally weighted partial least squares algorithm(LW-PLS)is the commonly used modeling method in just-in-time learning algorithms,and the selection of similarity samples is the most critical factor for LW-PLS.Many experimental studies have proved that the instant leaming algorithm based on regression coefficients and correlation coefficients to weight variables can significantly improve the prediction accuracy of the model,and the determination of variable weights will also affect the prediction effect.Therefore,based on the output-based just-in-time learning algorithm,the effects of different weight functions and different orders of weight coefficients on the modeling effect were explored,and verified in numerical examples and actual examples of sulfur recovery units.The results proved that in the actual industrial process,a reasonable selection of the weight function and a suitable number of weights can significantly improve the predictive performance of the system.