结合时变状态回忆匹配的常减压精馏过程自适应动态建模方法研究

Time-varying conditions in industrial processes have a significant impact on the performance of the data-driven dynamic model. The research project will expand research on dynamic modeling methods for product quality parameters in atmospheric and vacuum distillation processes, explore a new adaptive dynamic modeling method based on state-memories. Extraction the time-varying conditions from massive historical data as a typical state storage, which has the aspect of the memory feature in the human brain. Research on the similarity determine algorithm between current and historical status, which has the aspect of the searching and matching features in the human brain. It combines the dynamic modeling method with the moving window kernel learning algorithms and full uses of historical data, current data and future data to train the model. The improved intelligent optimization algorithms are used to choice the parameters for the dynamic model. In this way, it improves the ability of the dynamic model to deal with the time-varying condition. The research project will analyze the correlation about the atmospheric and vacuum distillation process parameters with the product quality parameters, explore the correlation between the input and output variables under the time-varying condition, to achieve structural updated of the dynamic model. Samples collected in the distillation process and the nominal data set to validate the effectiveness of the new method. The research project will provide a new method for intelligent dynamic modeling of critical parameters in the atmospheric and vacuum distillation process quality parameter estimation and provide the new ideas and theoretical basis for intelligent detection technology in industrial processes.

时变工况对基于数据驱动的工业过程动态模型性能有重大的影响。本项目将围绕时变工况条件下常减压精馏过程产品质量参数的动态建模相关问题展开研究，探索基于状态回忆的动态自适应建模新方法。模拟人脑在记忆方面的功能，研究特征工况提取算法，将海量历史数据中的时变工况作为典型状态存储；模拟人脑在回忆决策方面的功能，研究时变状态的相似性匹配算法，并将其和基于智能参数优化的移动窗核学习建模算法相结合，达到充分利用历史数据、当前数据和未来数据训练模型的目的，进而提高新方法处理时变问题的能力。针对常减压精馏过程产品质量参数和大规模模型输入变量的相关性，研究时变工况下模型输入与输出变量分析方法，实现模型的结构更新。通过收集的样本集和公开的标称数据集对新方法的有效性进行验证。本项目的研究将为常减压精馏过程质量参数的智能动态建模提供新的建模方法，为智能检测技术在工业过程关键参数估计中的应用提供新思路和理论依据。

常减压精馏过程产品质量关键参数软测量方法研究对于实现精馏生产过程参数检测、先进控制、节能增效等具有重要的意义。本项目针对多变量、工况时变条件下的石化企业常减压精馏过程产品质量关键参数软测量问题，从时变过程智能软测量的角度探索和研究了数据预处理、软测量辅助变量选择、基于核方法的模型泛化性、自适应动态校正等软测量方法中的若干关键难题，取得了相应的研究进展。在数据预处理方面，研究实现了基于Penalized阈值的小波包阈值去噪算法等数据预处理算法；在辅助变量选择算法方面，研究实现了一系列基于互信息、基于弹性网、基于PLS等辅助变量选择算法；在基于核方法的模型泛化性方面，研究了基于多核LSSVM系列算法的稀疏性和泛化性；在软测量自适应动态校正方面，研究了基于移动窗、即时学习等动态软测量校正算法。研究结果表明，上述算法更能适应动态时变过程，可以提高多变量、工况时变条件下的石化企业常减压精馏过程产品质量关键参数软测量模型的泛化精度。研究成果已经发表研究论文11篇，申请国家发明专利1项。