Submitted to the Electrical & Electronics Engineering Programme in Partial Fulfillment of the Requirements

This report serves to give the details of the development of the project and the steps taken in the realization of the Modeling of Primary Reformer Tube Metal Temperature (TMT). Literature surveys on the general idea of reformer, ammonia production, failures of the reformer tubes and the modeling related to dynamic systems have been conducted and discussed throughout this report. Many thanks to Vishal Nanji Patel for the early works done regarding the literature review, the early processing of the data as well as the modeling of primary reformer TMT using Artificial Neural Network.

LIST OF ABBREVIATIONS

CHAPTER 1 INTRODUCTION

Background Study
Problem Statement
Scope of Study The scope of study involves

Two factors mainly affect the breaking mechanism of tubes which is the thermal stress of the tube wall and by the stress imposed from the operation under pressure. The temperature of the tube metal (also called tube wall) affects the tube life as when overheating occurs; this causes dramatic reduction in tube life. The creep strength (the resistance to creep damage) of a tube is also affected by the tube design, tube material and the catalyst effects, apart from the tube wall temperature mentioned earlier.

CHAPTER 2

LITERATURE REVIEW/THEORY

Tube Failures

Modeling, Simulation, and Sensitivity Analysis of Steam Methane Reformers [81
Multiple Linear Regression (MLR)

The performance of neural networks is affected by some parameters such as neural network structure and the quality of the data preprocessing. In the implementation of control strategy, the neural network model is used to predict future plant performance. In this project, the neural network model will be trained only offline (batch) using real-time data obtained earlier from PASB.

Figure 6 : 100 % Tube Failure (left) and Tube Rupture (right)

CHAPTER 3 METHODOLOGY

Artificial Neural Network (ANN) Model Development First Phase

The correlation coefficient is basically the ratio of covariance to the product of standard deviations of two variables. If there is no relationship between two variables, the value of the correlation coefficient will indicate the zero value [4]. Thousands of iterations are performed to see the results of changing certain parameters that affect the performance of the neural network.

TMT modeling using the AutoRegressive Exogenous (ARX) parametric model was done via MATLAB's System Identification Toolbox. The order of the ARX model is determined and the parameters can be obtained using the least square method that finds a set of coefficients that minimizes the squared error for the entire set of known data. Evaluating the performance of the model, if it is not good enough, the process is repeated starting from the parameter estimation.

The tools used in developing the black box model to predict TMT in this project are: .. i) Neural network toolbox ii) System identification toolbox iii) Statistical toolbox.

CHAPTER 4

RESULTS AND DISCUSSION

Data Compilation and Level Selection [41

Both datasets are assembled and preprocessed for use in model development. Process variables and TMT values used in the data set are average values (average temperature of 144 tubes - for each chamber). Since the data are taken for Level I and Level 3, the higher mean TMT values are selected to be used as scores.

Based on the observation, the average values of TMT in level 3 have higher temperature readings compared to level 1. Thus, the data (TMT) from level 3 will be used as outputs for the model (room 1 and room 2) which will to develop.

Figure 12 : Average TMT Distribution for Level 1 and 3

Inputs/Process Variables Selection

The ANN model is developed using MATLAB's Neural Network Toolbox (see APPENDIX D for programming code). The performance of an artificial neural network (ANN) model depends on some parameters; the structure and preprocessing of the data. In ANN model development, several approaches are performed heuristically to obtain the optimal structure of a neural network model and are mentioned above in the methods section.

Based on the validation set for the TMT Chamber 1 results, it is observed that Variable Learning Rate Backpropagation (trainingda) performs the best with the lowest validated RMSE while taking the maximum epoch (1500) before training stopped. The result for the Levenberg-Marquardt (trainlm) shows the worst performance with the highest validated RMSE. The maximum epoch is increased to 2000 to give enough time for the result to converge to the next steps.

Compared to the previous results, the degradation of performances can be seen for the training's models while overall improvements on the validation's models. The iteration of changing the number of neurons from I to 100 neurons for the hidden layer is performed and analyzed on each training algorithm mentioned. The results show that for each training algorithm, with the exception of the Variable Learning Rate Backpropagation (traingda and traingdx), and Levenberg-Marquardt (trainlm) have the same trend whereby as the number of neurons increases, the RMSE decreases until it reaches an optimum number of neurons. and then RMSE starts to increase again (Fig. 14).

The performances of the training algorithms with the optimum number of neurons in terms of RMSE of the validation set are displayed in Table 7 .

When evaluating ANN performance, the validation set performance is considered much more important than the training set performance. The optimum results until this method is performed, RMSE for room 1 is 4.6165 and room 2 is 4.3553 respectively for the validation set and 15-16 % improved accuracy can be observed from the beginning of development of the ANN Model. The plot for the ANN model for the estimation/training and validation set of room 1 is shown as Fig.

The ANN model (MLP and Elman) is repeatedly trained and simulated with the increasing number of neurons for different training algorithms. The lowest value of valid RMSE values is recorded with the corresponding number of neurons for each training algorithm. The number of neurons is changed (1 to 100 neurons) for each algorithm to see which gives the best RMSE and FIT values.

The numbers of neurons producing the lowest RMSEs for the validation set are not the same for all algorithms. The best algorithm to train the model in predicting TMT can be any of the algorithms, but it depends on the application, speed and amount of memory to be used [9]. The algorithm that produces slightly the best results in terms of lowest RMSE for the validation set is traincgb.

However, the number of neurons needed in the hidden layer (in the range of 1 to 100 neurons) is not consistent (not the same) for both the chamber 1 and chamber 2 models, namely 92 neurons for chamber 1 and 99 neurons for the chamber 1 models. room 2 . .

Figure 16 : Validation Error's Model Scatter Plot

For the algorithms excl. traingda, traingdx and trainlm result for the lowest RMSEs that can be achieved, consistent, which is by using 4 neurons. For example, even if two Elman networks, with the same weights and biases, and receive identical inputs at a given time step, their outputs may be different due to different feedback modes [6]. It can be designed with the "newrbe" function, and no number of neurons needs to be defined.

The parameters to be set in this RBF model are the spreading constant that will affect the bias of the first layer [6]. From the results above, the performance of the chamber 2 model is the lowest when the dispersion value is equal to 0.185. Chamber 2's TMT model using the dispersion constant 0.185 has the lowest RMSE for the validation set value with 8.1673, but the FIT is only 5%.

The ARX model is developed using the user-friendly graphical user interface of MATLAB's System Identification Toolbox (Fig. 19). The parameters are then estimated and imported into the MATLAB workspace to be re-modeled using 199 data pairs (100 for training and 99 for validation). For the sake of simplicity, ARX's first model is called the ARX210 and the second model is the ARX410.

RMSE and FIT are used to measure the performance of the model, to later be compared with the previous ANN model and the MLR model.

Table 16 : Models' Performances with 0.185 SPREAD constant I d

Overall performance results

Compared to previous works (V.N. Patel), the performance of the model in predicting TMT has improved significantly from the initial performance, where the RMSE (validation) is 5.463 and 5.233 for chamber 1 and chamber 2, respectively, to 4.4614 and 4.1744 in this project . The accuracy or efficiency of the primary reformer TMT prediction model could be further improved from time to time. Robust TMT primary reformer models can be integrated into the existing control system to implement an appropriate control strategy (future work) if the developed predictive models are good enough.

CHAPTER 5

Conclusions

A proper model is needed to predict the TMT in the reformer to avoid overheating of the reformer tubes that could lead to unplanned plant shutdowns and losses. TMT primary reformer modeling is an ongoing development process and many things can be done to further improve the predictive model. The modeling that has been done so far focuses more on improving the performance based on the structure of the model (especially ANN).

To adjust the structure of the ANN model, it is necessary to consider the possibility of using genetic algorithms to speed up the convergence process at global minima. The selection of process variables is also a very critical area that will determine the accuracy and viability of TMT primary reformer modeling. Another robust statistical analysis can be performed to investigate the relationship between variables, to select the most critical variables and eliminate redundancy.

The model of primary reformer TMT that has been developed is based on the average TMT of 144 tubes (each chamber) and not the individual tubes. For more accurate and reliable TMT values, individual modeling can be developed to predict each of the individual TMT, but it will consume a lot of time (in modeling) and requires higher memory capacity and processing speed for real-time/simulation application. More fresh data must be obtained to provide a comprehensive database for developing a robust model.

Data reliability is also a limitation in developing a robust model where measuring TMT using a pyrometer is not really an accurate and reliable method.

34;Neural network cartridges for time series data mining", Proceedings of Joint International Conference on Neural Networks.

APPENDICES

PROJECT GANNT CHART FYP I

PROJECT GANNT CHART FYP II

LIST OF PROCESS VARIABLES AND CORRELATION COEFFICIENTS

MLP OR ELM MATLAB CODING AND VALUES

RBF MATLAB CODING

ARX MATLAB CODING AND VALUES