有偏最小最大概率模型及在汽油属性预测中的应用

Biased minimax probability model and its application in prediction of gasoline properties

中文关键词  汽油调合  最小最大概率机  动态建模  机器学习  过程系统

英文关键词  gasoline blending  minimax probability machine  dynamic modeling  machine learning  process systems

中文摘要

      汽油属性的在线预测多采用无偏估计方法建立的近红外定量分析模型实现, 累积预测误差的正负偏差范 围难以控制, 这会严重影响汽油调合优化控制的投运效果. 针对这一问题, 本文提出了一种采用有偏估计实现油品 属性在线预测的方法. 首先从最小最大概率学习机出发, 提出了有偏最小最大概率回归模型. 然后利用即时学习方 法设计了有偏回归模型的局部建模与更新策略, 用以提高回归模型的自适应能力. 最后在国内某炼厂汽油调合过 程中采集的工业数据上进行实验, 结果表明该方法与传统方法相比具有明显优势, 有利于大幅度提高调合优化控制 的投运率.

英文摘要

      The online prediction of gasoline properties is mostly realized by the near-infrared quantitative analysis model which established by the unbiased estimation method. However, the range of the positive and negative deviations of the cumulative prediction error is difficult to control, which will seriously affect the operation of gasoline blending optimization control. To deal with this issue, a biased estimation method is proposed for the online prediction of gasoline properties. Firstly, a biased minimax probability regression model is proposed based on minimax probability machine. Then, based on just-in-time learning approach, a local modeling and updating strategy is developed for the biased regression model to improve its adaptive ability. Finally, experiments are carried out with the gasoline data collected from a domestic oil refinery. The results show that the present method has obvious advantages compared with traditional algorithms, and it is beneficial to improve the operation rate of the optimal control system of gasoline blending.