本研究以苯乙烯丙烯腈(SAN)高分子工廠為載體，由實廠數據分別建立熔融指數(MI)預測模型以及回收槽出口濃度估算模型。透過模型輸出每十分鐘一筆模擬結果提供現場人員兩變數參考值，以克服量測限制導致難以得知即時MI以及回收槽出口濃度的問題。
MI預測模型經數據分析選定變數並透過GRU模型輸出MI。以不同輸入變數建立場景切換模型及監控模型，輸入變數是否加入中介變數將影響模型功能與效能；其中，八變數監控模型測試集之MAPE為3.629 %。結合建立完成之模型與虛擬控制器計算TDDM注入量建議值，現場操作人員可參考建議值或操作經驗調整其流量，再由實際操作流量計算下一時刻之建議值。MI預測模型採用建議值作為輸入變數，將縮短品別轉換花費時間且MI較貼近設定點。模擬結果與實際量測MI品別轉換花費時間相比最長相差28.6小時，達減少原物料浪費與降低次級品產量之目的。
回收槽出口濃度估算模型為結合數據驅動與理論之複合模型，選定之輸入變數透過模型輸出回收槽進口濃度再由理論模型計算回收槽出口濃度。計算結果與實際量測濃度差異小於一個單位，偏差皆介於誤差範圍內。複合模型計算結果接著計算混合槽出口濃度，用以判斷進入反應器物流組成是否介於管制範圍內，計算結果之MAPE皆小於2%。模型每十分鐘輸出一筆混合槽出口濃度，若計算結果超出管制範圍內可及時調整混合槽單體流量，解決量測限制導致難以及時調整單體進料量之問題。
將MI預測模型與回收槽出口濃度模型實際應用於製程中，透過滾動的方式持續輸出結果，避免過多數據累積於系統中造成計算的負荷，最終模擬結果提供實廠變數監控值以及操作建議值。

In this study, the styrene acrylonitrile (SAN) polymer factory was used as the carrier, and the melt index prediction model and the recycling tank output concentration estimation model were established from the actual factory data. The models will output simulation results every ten minutes and provide reference values of two variables for on-site operators to overcome the measurement limitation issue.
Using data analysis selects input variable. Then, the MI prediction model output MI with GRU model. According to different input variables establish a scene switching model and monitoring models. The input variables including mediator variables will affect the function and performance of the model. Among them, the MAPE of the eight-variable monitoring model in testing set is 3.629%. Combining well-trained model with the virtual controller calculate the suggestion value of TDDM feed flowrate. On-site operators can adjust the feed flowrate with the suggestion value or operating experience, and using the actual operating flowrate calculate the suggestion value at the next moment. Taking the suggestion value as the input variable of MI prediction model, it will shorten the time taken for grade transition and the MI will be closer to the set point. The simulation result can shorten the grade transition time up to 28.6 hours. It can achieve the purpose of reducing the waste of raw materials and decreasing the yield of substandard products.
The recycling tank output concentration estimation model is a hybrid model that combine data driven model and theoretical model. Using selected input variables output recycling tank input concentration and calculate recycling tank output concentration with theoretical model. The error between calculation result and actual measure concentration is within one unit and the deviation is within the control range. Then, the calculation result of the hybrid model can be used to calculate the mixing tank output concentration and determine whether the composition of the reactor feed flowrate is within the control range. The calculated results of MAPE are all less than 2%. The calculation result provides the mixing tank output concentration every ten minutes. If the calculation result is out of the control range, the monomer feed flowrate of the mixing tank can be adjusted immediately. The problem of hard to adjust the monomer feed flowrate in time can be solved.
The MI prediction model and the recycling tank output concentration model are actually applied in the actual process, and the results are continuously output through a rolling method to avoid the data accumulate in the system. The final simulation results provide real plant variable monitoring values and operating suggestions.

[英文]
1. Åström, K. J., & Hägglund, T. (1984). Automatic tuning of simple regulators with specifications on phase and amplitude margins. Automatica, 20(5), 645-651.
2. Basheer, I. A., & Hajmeer, M. (2000). Artificial neural networks: fundamentals, computing, design, and application. Journal of microbiological methods, 43(1), 3-31.
3. Bergstra, J., & Bengio, Y. (2012). Random search for hyper-parameter optimization. Journal of machine learning research, 13(2).
4. Cho, K., Van Merriënboer, B., Gulcehre, C., Bahdanau, D., Bougares, F., Schwenk, H., & Bengio, Y. (2014). Learning phrase representations using RNN encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078.
5. Christopher, O. (2015). Understanding LSTM Networks, Retrieved from http://colah.github.io/posts/2015-08-Understanding-LSTMs/ (May 16, 2021)
6. Fernandes, F. A. N., & Lona, L. M. F. (2005). Neural network applications in polymerization processes. Brazilian Journal of Chemical Engineering, 22(3), 401-418.
7. Han, I. S., Han, C., & Chung, C. B. (2005). Melt index modeling with support vector machines, partial least squares, and artificial neural networks. Journal of Applied Polymer Science, 95(4), 967-974.
8. Jumari, N. F., & Yusof, K. M. (2017). Comparison of Melt Flow Index of Propylene Polymerisation in Loop Reactors using First Principles and Artificial Neural Network Models. Chemical Engineering Transactions, 56, 163-168.
9. Kumar, Munna & Tiwari, Awadhesh & Singh, R S. (2015). A comparative study on Tuning of PID controller. Paper presented at the 68th Annual Session of the Indian Institute of Chemical Engineers, Guwahati
10. Mikko Tahkola (2019), Developing Dynamic Machine Learning Surrogate Models of Physics-Based Industrial Process Simulation Models, University of Oulu Faculty of Technology, Finland.
11. Noor, R. M., Ahmad, Z., Don, M. M., & Uzir, M. H. (2010). Modelling and control of different types of polymerization processes using neural networks technique: A review. The Canadian journal of chemical engineering, 88(6), 1065-1084.
12. Schmidhuber, J., & Hochreiter, S. (1997). Long short-term memory. Neural Comput, 9(8), 1735-1780.
13. Seborg, D. E., Mellichamp, D. A., Edgar, T. F., & Doyle III, F. J. (2010). Process dynamics and control. John Wiley & Sons.
14. Seong, A. R., Lee, E. H., Lee, K. N., & Yeo, Y. K. (2009). Modeling and control of the melt index in HDPE process. Korean Journal of Chemical Engineering, 26(1), 7-13.
15. Shahriari, B., Swersky, K., Wang, Z., Adams, R. P., & De Freitas, N. (2015). Taking the human out of the loop: A review of Bayesian optimization. Proceedings of the IEEE, 104(1), 148-175.
16. Sharma, S., & Sharma, S. (2017). Activation functions in neural networks. Towards Data Science, 6(12), 310-316.
17. Shewalkar, A. (2019). Performance evaluation of deep neural networks applied to speech recognition: RNN, LSTM and GRU. Journal of Artificial Intelligence and Soft Computing Research, 9(4), 235-245.
18. Sun, Q., & Ge, Z. (2021). A survey on deep learning for data-driven soft sensors. IEEE Transactions on Industrial Informatics.
19. Yu, T., & Zhu, H. (2020). Hyper-parameter optimization: A review of algorithms and applications. arXiv preprint arXiv:2003.05689.
[中文]
1. Tom (2018)。資料的正規化與標準化。檢自：https://aifreeblog.herokuapp.com/posts/54/data_science_203/ (May 03, 2021)
2. 刘金龙(2016)。SAN 树脂工艺技术浅析。橡塑资源利用，04，20~24
3. 周秉誼(2016)。淺談Deep Learning原理及應用。臺灣大學計資中心電子報，38(期)
4. 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，台灣科技大學