Modeling complex process systems using the recurrent multilayer perceptron /
The objective of this research study is to develop a method for designing Computational Neural Networks (CNNs)-based empirical models capable of performing accurate multi-step-ahead (SSP) and single-step-ahead prediction (MSP) for complex process systems. Empirical models which can perform accurate...
| Main Author: | |
|---|---|
| Format: | Thesis Book |
| Language: | English |
| Published: |
[Place of publication not identified] :
[publisher not identified] ;
1995.
|
| Subjects: | |
| Online Access: | Link to OAKTrust copy http://proxy.library.tamu.edu/login?url=http://proquest.umi.com/pqdweb?did=731677271&sid=1&Fmt=2&clientId=2945&RQT=309&VName=PQD |
| Summary: | The objective of this research study is to develop a method for designing Computational Neural Networks (CNNs)-based empirical models capable of performing accurate multi-step-ahead (SSP) and single-step-ahead prediction (MSP) for complex process systems. Empirical models which can perform accurate MSP have many applications in the areas of forecasting and prediction, control engineering, and condition monitoring and fault diagnosis. Conventional modeling methods utilizing first principles have many disadvantages, such as big development cost, potential inaccuracies in view of drifts, inability to incorporate tear and wear elects, to just name a few. Moreover, for complex process systems the conventional techniques of System Identification (SI) fail to give models capable of performing accurate MSP. The proposed empirical modeling method snakes use of the Recurrent Multilayer Perception (RMLP) neural network. The RMLP is a hybrid, feedforward and feedback, neural network which exhibits local information feedback through time-delayed recurrence and cross-talk. Two learning algorithms are considered when using the RMLP, the Teacher Forcing (TF) algorithm, and the Global Feedback (GF) algorithm. Both are based on a dynamic gradient descent approach. The difference between the TF and the GF algorithms is that the former utilizes past sensed systems outputs, whereas the latter utilizes past RMLP predictions as inputs to the network. The proposed method also includes guidelines for choosing the appropriate network architecture and the training stopping criteria. An artificial case-study, as well as a U-Tube Steam Generator (UTSG), were used to validate the accuracy of the proposed modeling method. For the artificial problem, the model based on conventional identification methods gave good results, while the CNN-based model gave slightly improved results. However, both models were capable of performing accurate MSP. For the UTSG, the CNN models performed well in MSP as well as in SSP. Validation studies using the developed CNN models have demonstrated that they exhibit substantial generalization of the operational UTSG dynamics. Models based on conventional identification methods failed to achieve such results. The success of the proposed modeling method is believed to be the results of the GF used in the RMLP which provides it with global and local memory, the associated dynamic learning algorithm, and the guidelines used in choosing the network achitecture. The prediction results obtained when using TF in the RMLP were much better than the results obtained using conventional SI algorithms. However, these results were still inferior compared to the results obtained using GF in the RMLP. This research study and the accompanying results demonstrate that the proposed CNN design method and the associated algorithms provide a feasible and robust procedure for modeling complex process systems. |
|---|---|
| Item Description: | Vita. "Major Subject: Nuclear Engineering". |
| Physical Description: | xv, 200 leaves : illustrations ; 28 cm. Issued also on microfiche from University Microfilm Inc. |
| Bibliography: | Includes bibliographical references (leaves 192-197). |