Title Page
Contents
NOMENCLATURES 11
ABBREVIATION 12
ABSTRACT 13
I. INTRODUCTION 15
II. PROCESS DESCRIPTION 17
2.1. Introduction of CFB Boiler 17
2.2. CFB Boiler Concept 20
2.3. CFB Boiler Configuration 21
2.4. Features of CFB Boiler 25
2.4.1. Structure of CFB Boiler 25
2.4.2. Advantages of CFB Boiler 26
2.5. Emissions from the Boiler 33
2.5.1. Air Pollution 33
2.5.2. Sulfur Dioxide Emission 36
2.5.3. Nitrogen Oxide Emission 44
2.5.4. Carbon Emissions 48
III. METHODOLOGY 49
3.1. Introduction of Machine Learning 49
3.1.1. The Task, T 49
3.1.2. The Performance Measure, P 50
3.1.3. The Experience, E 50
3.2. Support Vector Machine 54
3.2.1. Linear SVM Classification 54
3.2.2. Nonlinear SVM Classification 55
3.2.3. SVM Regression 56
3.2.4. SVM Theory 58
3.3. Deep Neural Network 61
3.3.1. Perceptron 61
3.3.2. Multi-Layer Perceptron and Backpropagation 64
3.3.3. Regression MLP 67
3.3.4. Fine-Tuning Neural Network Hyperparameters 68
3.4. Convolutional Neural Network 71
3.4.1. Convolutional Layer 71
IV. PROBLEM FORMULATION 79
4.1. Target Process Description 79
4.2. Machine Learning Strategy 80
4.2.1. Data Preprocess 81
4.2.2. Feature Selection 81
4.2.3. Model Train 82
4.2.4. Validation 84
V. RESULTS AND DISCUSSION 85
5.1. Data Preprocess 85
5.2. Feature Selection 88
5.3. Model Train 94
5.3.1. SVM Model Train 94
5.3.2. DNN Model Train 100
5.3.3. CNN Model Train 106
5.4. Validation 112
VI. CONCLUSION AND FUTURE WORK 125
VII. REFERENCE 126
국문요약 131
Table 2.1. Comparison of CFB with other types of boilers 19
Table 5.1. The performance results of the trained SVM algorithms for NOx-SOx 99
Table 5.2. The performance results of the trained DNN algorithms for NOx-SOx 105
Table 5.3. The performance results of the trained CNN algorithms for NOx-SOx 111
Figure 2.1. Schematic diagram of a CFB boiler 17
Figure 2.2. Comparison of design characteristics of different types of firing for boilers 19
Figure 2.3. Components of a typical circulated fluidized bed boiler 21
Figure 2.4. INTREX heat exchanger 23
Figure 2.5. Comparison of the grate heat-release rate of different firing systems 30
Figure 2.6. Absorption of sulfur dioxide by sorbents 38
Figure 2.7. Effect of combustion temperature on sulfur capture efficiency. 41
Figure 2.8. Limestone demand decreases with the gas residence time for a specific level of sulfur capture 42
Figure 2.9. Relative importance of different reduction paths in the formation and reduction of nitric oxide is indicated through numerical figures. 46
Figure 3.1. Supervised learning algorithms 51
Figure 3.2. Unsupervised learning algorithms 52
Figure 3.3. Reinforcement learning algorithms 53
Figure 3.4. Large margin classification 54
Figure 3.5. Hard margin sensitivity to outliers 55
Figure 3.6. Adding features to make a dataset linearly separable 56
Figure 3.7. SVM regression 57
Figure 3.8. SVM regression using a 2nd-degree polynomial kernel 57
Figure 3.9. Decision function for an example dataset 59
Figure 3.10. A smaller weight vector results in a larger margin 60
Figure 3.11. Threshold logic unit 61
Figure 3.12. Perceptron diagram 62
Figure 3.13. XOR classification problem and an MLP that solves it 64
Figure 3.14. Multi-Layer perceptron 65
Figure 3.15. Activation functions and their derivatives 67
Figure 3.16. CNN layers with rectangular local receptive fields 71
Figure 3.17. Connections between layers and zero padding 72
Figure 3.18. Reducing dimensionality using a stride of 1 73
Figure 3.19. Applying two different filters to get two feature maps 74
Figure 3.20. Max pooling layer (2x2 padding kernel, stride 2, no padding) 76
Figure 3.21. Invariance to small translations 77
Figure 3.22. Typical CNN architecture 78
Figure 4.1. CFB boiler process in SAMCHEOK coal power plant 79
Figure 4.2. Layout of the machine learning sequence for building prediction models 80
Figure 4.3. SVM regression algorithm training procedure 83
Figure 4.4. Deep Neural Network algorithm training procedure 84
Figure 4.5. Convolutional Neural Network algorithm training procedure 84
Figure 5.1. NOx data preprocess 86
Figure 5.2. SOx data preprocess 87
Figure 5.3. NOx data distribution plot 89
Figure 5.4. SOx data distribution plot 90
Figure 5.5. NOx data heatmap plot 91
Figure 5.6. SOx data heatmap plot 92
Figure 5.7. Data partitioning strategy employed on the prepared dataset for the models' learning 93
Figure 5.8. Training performance of the SVM for NOx 95
Figure 5.9. Training performance of the SVM for SOx 96
Figure 5.10. Overall performance of the SVM for NOx emission prediction 97
Figure 5.11. Overall performance of the SVM for SOx emission prediction 98
Figure 5.12. The prediction results of the trained SVM algorithms for NOx-SOx 99
Figure 5.13. Training performance of the DNN for NOx 101
Figure 5.14. Training performance of the DNN for SOx 102
Figure 5.15. Overall performance of the DNN for NOx emission prediction 103
Figure 5.16. Overall performance of the DNN for SOx emission prediction 104
Figure 5.17. The prediction results of the trained DNN algorithms for NOx-SOx 105
Figure 5.18. Training performance of the CNN for NOx 107
Figure 5.19. Training performance of the CNN for SOx 108
Figure 5.20. Overall performance of the CNN for NOx emission prediction 109
Figure 5.21. Overall performance of the CNN for SOx emission prediction 110
Figure 5.22. The prediction results of the trained DNN algorithms for NOx-SOx 111
Figure 5.23. The applied algorithms results for predicting NOx A) NOx-SVM, B) NOx-DNN, C) NOx-CNN 113
Figure 5.24. The applied algorithms results for predicting SOx A) SOx-SVM, B) SOx-DNN, C) SOx-CNN 114
Figure 5.25. The comparison about accuracy for NOx-SOx A) Accuracy for NOx, B) Accuracy for SOx 115
Figure 5.26. The comparison about MAE for NOx-SOx A) MAE for NOx, B) MAE for SOx 117
Figure 5.27. The comparison about RMSE for NOx-SOx A) RMSE for NOx, B) RMSE for SOx 119
Figure 5.28. The comparison about MAPE for NOx-SOx A) MAPE for NOx, B) MAPE for SOx 121
Figure 5.29. The comparison about R2 for NOx-SOx A) R2 for NOx, B) R2 for SOx 123