化工製程為因應終端生產需求，故對於產品性質管控上具有一定的標準。然
而，產品性質之線上量測往往需付出昂貴成本或甚至無法線上進行，通常透過人
工間隔數小時進行採樣，因此易造成產品性質採樣頻率過低，而大幅增加品質管
控上的困難度。此外，許多化工製程生產中時常面臨產品規格間之轉換，或者生
產時存在進料條件、產能變化等因素之干擾。然而，於上述情境中，許多用於品
質控制之操作變數無明確之調整依據，通常以現場人員經驗進行調整，導致產品
規格轉換時間過長，造成原物料浪費，或產生發生品質震盪之問題。
針對產品性質測量產生之問題，本研究使用非線性循環類神經網路模式
(Recurrent Neural Network, RNN)於產品性質之動態預測，實現對品質之實時監控
與未來趨勢之預測，進而提升品質管控效果。針對產品規格切換或干擾排除之問
題，本研究利用 RNN 模式設計虛擬 PID 控制器及 RNN 模式預測控制控制器
(RNN-based model predictive control, RNN-MPC)兩種控制算法，實現製程導航技
術，用以實時提供現場人員操作變數之調控指引，改善品質震盪或產品規格轉換
時間過長之問題。為確保製程導航系統之正確性，本研究針對RNN 模式之程序
增益(Process gain)方向性進行研究，並考慮實際程序與模式間的程序增益一致性
(Gain consistency)，合理的設計類神經網路模式於參數估計時之目標函數，使建
模同時考慮準確性及實際物理現象之約束，避免發生操作指引方向性錯誤問題。
本研究以實際高分子製程為案例，其中高分子產品之熔融指數(Melt index,
MI)為品質指標，而斷鏈劑(Chain modifier)添加量為品質控制之操作變數。因此，
將非線性 RNN 模式用於 MI 預測，製程導航系統用於斷鏈劑添加操作指引。結
果顯示，RNN 之 MI 模擬測試結果之平均絕對百分誤差(MAPE)為 5.4 %，且於不
同測試條件下之增益一致性為 100 %；製程導航系統之自動控制與實廠手動控制
策略相比，於各品別操作條件下，MI 與目標值之積分絕對誤差(IAE)平均約可下
降80 %，且於不同品別轉換情境下之轉換時間平均約可縮短約 31 小時。

The chemical process is in response to the end product needs, so there are standards for the control of product properties. However, online measurement of product properties often requires expensive costs or even cannot be performed online. Usually, sampling is performed manually at intervals of several hours. Therefore, it causes low sampling frequency of product properties, which greatly increases the difficulty of quality control. In addition, the production of many chemical processes often encounters the transition of product specifications or the disturbance, such as feed conditions and production rate changes during production. However, in the above situations, many manipulated variables used for quality control have no clear basis for adjustment. They are usually adjusted based on the experience of plant operator, resulting in long product transition time, waste of raw materials, or quality fluctuation when encountering disturbances.
Aiming at the problems arising from the measurement of product properties, this study uses the nonlinear recurrent neural network (RNN) model to dynamically predict product properties to realize real-time monitoring of quality and prediction of future trends, thereby improving the effectiveness of quality control. This work developed a virtual PID controller, and model predictive control based on the proposed RNN model, in order to deal with the operation problems (i.e. quality fluctuations and long transition time). To ensure the correctness of the process guidance system, this research studies the process gain direction of the RNN model, considers the process gain consistency between the actual process and the model, and reasonably designs the objective function in parameter estimation. This enables the training to consider the constraints of accuracy and actual physical constraints (i.e. process gain sign) at the same time, as well as avoiding the problem of misleading operation guidance.
This study takes the actual polymer process as an example, in which the melt index (MI) of polymer products is the quality index, and the chain modifier feed flow rate is a manipulated variable for quality control. Therefore, the nonlinear RNN model was used for MI prediction, and the process guidance system was used for chain modifier feed operation guidance. The results show that the average absolute percentage error (MAPE) of the MI simulation test results of RNN is 5.4 %, and the gain consistency under different test conditions is 100 %. The automatic control of the process guidance system is compared with the actual plant manual control strategy. Under the operating conditions of different grades, the integral absolute error (IAE) between MI and the target value can be reduced by about 80 %, and the transition time can be shortened by about 31 hours significantly in the different grade transition situations.

[英文]
1. Chen, T., & Guestrin, C. (2016, August). Xgboost: A scalable tree boosting
system. In Proceedings of the 22nd acm sigkdd international conference on
knowledge discovery and data mining (pp. 785-794).
2. Chien, I. L., Kan, T. W., & Chen, B. S. (2007). Dynamic simulation and operation
of a high pressure ethylene-vinyl acetate (EVA) copolymerization autoclave
reactor. Computers & chemical engineering, 31(3), 233-245.
3. Chou, C. H., Wu, H., Kang, J. L., Wong, D. S. H., Yao, Y., Chuang, Y. C., ... &
Ou, J. D. Y. (2019). Physically consistent soft-sensor development using
sequence-to-sequence neural networks. IEEE Transactions on Industrial
Informatics, 16(4), 2829-2838.
4. Chung, J., Gulcehre, C., Cho, K., & Bengio, Y. (2014). Empirical evaluation of
gated recurrent neural networks on sequence modeling. arXiv preprint
arXiv:1412.3555.
5. Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural
computation, 9(8), 1735-1780.
6. Hsiao, Y. D., Kang, J. L., & Wong, D. S. H. (2021). Development of robust and
physically interpretable soft sensor for industrial distillation column using transfer
learning with small datasets. Processes, 9(4), 667.
7. Ke, W., Huang, D., Yang, F., & Jiang, Y. (2017, November). Soft sensor
development and applications based on LSTM in deep neural networks. In 2017
IEEE Symposium Series on Computational Intelligence (SSCI) (pp. 1-6). IEEE.
8. Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic
optimization. arXiv preprint arXiv:1412.6980.
9. Lee, H. Y., Yang, T. H., Chien, I. L., & Huang, H. P. (2009). Grade transition
using dynamic neural networks for an industrial high-pressure ethylene–vinyl
acetate (EVA) copolymerization process. Computers & Chemical
Engineering, 33(8), 1371-1378.
10. Li, M. (2016). A Tutorial On Backward Propagation Through Time (BPTT) In
The Gated Recurrent Unit (GRU) RNN. Dept. Comput. Sci., Univ. British
Columbia, Vancouver, BC, Canada, Tech. Rep.
11. Liu, L., Jiang, H., He, P., Chen, W., Liu, X., Gao, J., & Han, J. (2019). On the
variance of the adaptive learning rate and beyond. arXiv preprint
arXiv:1908.03265.
12. Luyben, M. L., Tyreus, B. D., & Luyben, W. L. (1997). Plantwide control design
procedure. AICHE journal, 43(12), 3161-3174.
13. Powell, M. J. (1994). A direct search optimization method that models the
objective and constraint functions by linear interpolation. In Advances in
optimization and numerical analysis (pp. 51-67). Springer, Dordrecht.
14. Qin, S. J., & Badgwell, T. A. (2003). A survey of industrial model predictive
control technology. Control engineering practice, 11(7), 733-764.
15. Tian, Y., Zhang, J., & Morris, J. (2001). Modeling and optimal control of a batch
polymerization reactor using a hybrid stacked recurrent neural network
model. Industrial & engineering chemistry research, 40(21), 4525-4535.
16. Xia, L. Y., Pan, H. T., Zhou, M. F., Cai, Y. J., & Sun, X. F. (2011). Modeling
Methodology Based on Stacked Neural Networks for Inferential Prediction of
Polypropylene Melt Index. In Applied Mechanics and Materials (Vol. 55, pp. 670-
674). Trans Tech Publications Ltd.
17. Zhang, J. (2005). Batch process modeling and optimal control based on neural
network models. Acta Automatica Sinica, 31(1), 19-31.
18. Zhang, J. (2008). Batch-to-batch optimal control of a batch polymerisation process
based on stacked neural network models. Chemical Engineering Science, 63(5),
1273-1281.
19. Zhang, M., Lucas, J., Ba, J., & Hinton, G. E. (2019). Lookahead optimizer: k steps
forward, 1 step back. Advances in neural information processing systems, 32.
20. Ziegler, J. G., & Nichols, N. B. (1942). Optimum settings for automatic
controllers. trans. ASME, 64(11).
[中文]
1. 李豪業(2000)。非線性不穩定程序之多重線性模式預測控制，碩士論文，長庚
大學
2. 李豪業(2018)。虛實系統的概念與化工製程上的應用，化工，第65 卷第2 期。
3. 羅宇軒(2019)。類神經網路模型應用於實廠 SAN 高分子製程之研究，碩士論
文，臺灣科技大學
4. 陳建穎(2020)。使用類神經網路及模式預測控制於乙酸異丙酯反應蒸餾製程
之控制研究，碩士論文，臺灣科技大學
5. 王嘉懿(2020)。長短期記憶模型應用於實廠聚丙烯製程熔融指數與乙烯含量
預測之研究，碩士論文，臺灣科技大學
6. 張君詠(2021)。應用類神經網路及理論模型於 SAN 高分子品別轉換之研究，
碩士論文，臺灣科技大學
7. 黃敬偉(2021)。試驗廠級乙酸異丙酯反應蒸餾製程之類神經網路模型建置與
干擾排除指引研究，碩士論文，臺灣科技大學