基於模型的大規模系統數據校正與參數估計

Model based DRPE large scale system

莊英譽(Ying-Yu Chuang)

指導教授 ： 陳榮輝

中原大學/工學院/化學工程研究所/碩士(2015年)

https://doi.org/10.6840/CYCU.2015.00399

摘要

數據校正與參數估計是同時考慮可靠的校正數據，並對模型參數進行估計。過去的文獻中大都是基於簡化的機理模型進行研究，但如果是大規模複雜程序系統，受到計算機性能限制使求解困難。此外，程序的操作條件是在大範圍切換，模型參數呈非線性變化，使得數據校正與參數估計更加困難。本研究提出並行運算技術，能有效處理大規模程序系統受到計算機限制的問題。由於並行運算是採用多個處理器負擔系統的運算，使得系統能夠求解。當程序是多操作條件下，提出聚類技術，形成數個局部的模型，在提升計算效能同時依然能夠提升數據校正與參數估計準確度。為了減少多操作條件系統，因為初始值導致求解失敗，提出即時學習法，定義以Bayesian機率找到相似操作條件之相似因子分析，即使初始值差，也能夠提升求解效率及性能。為了驗證上述方法的有效性，分別應用於高分子聚合反應程序系統和空氣分離反應程序系統。

關鍵字

大規模系統 ； 數據校正與參數估計

並列摘要

Data reconciliation and parameter estimation (DRPE) are important tasks for improving the performance of the process design and analysis. They provide reliable, reconciled values of measurements and estimated values of parameters to make sure that the identified process models can accurately describe the behavior of the industrial process. Since the difficulty of solving nonlinear DRPE optimization problems increases significantly with the growing number of variables and the multi-operating conditions, particularly in a large-scale process system, solving many smaller DRPE sub-problems iteratively can be more efficient than solving the whole large scale DRPE problem. With the features of parallel processing technology, a novel approach is proposed to find the optimal distributed DRPE sub-problems in this research work. To solve the optimal decomposition optimization sub-problems, the clustering based logical equation set decomposition (CLESD) is developed to reduce the sizes of sub-DRPE problems and to minimize loss information of the original large DRPE problem. The decomposition decomposes a large DRPE problem in two stages, including the variable clustering decomposition as well as the equation clustering decomposition. Also, a proper definition of a decomposition index and an efficient analysis algorithm are required. For the optimization problems with multi-operating conditions, a methodology with the combination of similar operating conditions, including the steady state detection and the clustering of multi-operating conditions, is proposed to construct a smaller DRPE problem. Further, the initial guesses influence the result of optimization problems. An efficient method, just-in-time learning-based DRPE (JITL-DRPE), is proposed to improve the efficiency of computing. The Bayesian approach is used to construct the similarity factor for finding appropriate initial guesses. Finally, we compare the traditional DRPE method with the proposed DRPE method through three industrial applications. The results show that the proposed method outperforms the traditional DRPE method in terms of the solution time and the number of iterations.

並列關鍵字

large scale system ； data reconciliation and parameter estimation

參考文獻

[1] Albuquerque, J.S. ; L.T. Biegler, Data reconciliation and gross‐error detection fordynamic systems, AIChE J 42 (1996) 2841–2856.

[2] Arora, N. ; L.T. Biegler, Redescending estimators for data reconciliation and parameter estimation, Comput. Chem. Eng. 25 (2001) 1585–1599.

[3] Chen, Z. ; X. Chen, Z. Shao, Parallel calculation methods for molecular weight distribution of batch free radical polymerization, Comput. Chem. Eng. 48 (2013) 175–186.

[4] Chen, J. ; Y. Peng, J.C. Munoz, Correntropy estimator for data reconciliation, Chem.Eng. Sci. 104 (2013) 1019–1027.

[5] Daszykowski, M. ; K. Kaczmarek, Y. Vander Heyden, Robust statistics in data analysis—a review: basic concepts, Chemom. Intell. Lab. Syst. 85 (2007) 203–219.