표제지
목차
국문초록 15
Abstract 17
제1장 서론 19
1.1. 연구의 배경 19
1.2. 연구의 목적 21
1.3. 연구의 방법 및 내용 23
제2장 머신러닝의 이론적 배경 26
2.1. 머신러닝(Machine Learning)의 개념 26
2.2. 머신러닝에 사용된 모델 29
2.2.1. 선형회귀모델(Linear Regression Model) 29
2.2.2. 규제 선형회귀모델(Regularized Linear Regression Model) 31
2.2.3. 후버회귀모델(Huber Regression Model) 32
2.2.4. 란삭회귀모델(RANSAC Regression Model) 33
2.2.5. 서포트 벡터 머신(Support Vector Machine) 34
2.2.6. 결정 트리(Decision Tree) 36
2.2.7. 랜덤 포레스트(Random Forest) 38
2.3. 인공신경망(Artificial Neural Network, ANN) 40
제3장 기존 쇄파지표 경험식의 특성 49
3.1. 쇄파에 대한 물리량 정의 49
3.2. 쇄파 예측을 위한 기존 경험식의 특성 51
3.2.1. 쇄파에 대한 경험식 고찰 51
3.2.2. 쇄파지표 예측을 위한 경험식 분류 54
3.3. 기존의 쇄파 실험결과와의 비교 56
3.3.1. 실험자료 수집 56
3.3.2. 쇄파지표에 대한 목표변수 설정 57
3.3.3. 기존 실험결과와 경험식의 비교 62
제4장 머신러닝에 기반한 쇄파지표 개발 74
4.1. 머신러닝 모형 수립 74
4.1.1. 입력변수 및 교차검증 74
4.1.2. 하이퍼 파라미터 최적화 75
4.1.3. 머신러닝 API 적용 76
4.2. 머신러닝 모델별 쇄파지표 예측성능 78
4.2.1. 선형회귀모델(Linear Regression Model, LRM) 78
4.2.2. 릿지회귀모델(Ridge Regression, RR) 82
4.2.3. 후버회귀모델(Huber Regression, HR) 90
4.2.4. 란삭회귀모델(RANSAC Regression) 100
4.2.5. 서포트벡터머신(Support Vector Machine, SVM) 110
4.3. 쇄파지표식 제안 120
4.3.1. 알고리즘별 회귀계수 선정 120
4.3.2. 쇄파지표식 제안 122
제5장 RF와 ANN 알고리즘을 이용한 쇄파예측모델 130
5.1. 랜덤포레스트(Random Forest, RF) 130
5.1.1. 매개변수 설정 130
5.1.2. 하이퍼 파라미터 131
5.1.3. 쇄파지표의 예측결과 132
5.2. 인공신경망(Artificial Neural Network, ANN) 145
5.2.1. 입력변수 정규화 및 훈련 자료 결정 145
5.2.2. 은닉층 및 노드의 개수 설정 145
5.2.3. 쇄파지표의 예측결과 146
5.3. 머신러닝과 경험식과의 예측결과 비교 152
제6장 요약 및 결론 155
참고문헌 159
Table 3.1. Summary of historical experimental data set on breaking waves 56
Table 3.2. MER, SI and R² of wave breaking height prediction 70
Table 3.3. Erms, Ebias and Esd of wave breaking height prediction[이미지참조] 71
Table 3.4. MER, SI and R² of wave breaking depth prediction 73
Table 3.5. Erms, Ebias and Esd of wave breaking depth prediction[이미지참조] 73
Table 4.1. LRM prediction results for breaking wave height 80
Table 4.2. LRM prediction results for breaking wave depth 81
Table 4.3. RR prediction results for breaking wave height 84
Table 4.4. RR prediction results for breaking wave depth 87
Table 4.5. HR prediction results for breaking wave height 91
Table 4.6. HR prediction results for breaking wave depth 96
Table 4.7. RANSAC prediction results for breaking wave height 101
Table 4.8. RANSAC prediction results for breaking wave depth 106
Table 4.9. SVM prediction results for breaking wave height 112
Table 4.10. SVM prediction results for breaking wave depth 116
Table 4.11. Summary of ML prediction results for breaking wave height 120
Table 4.12. Summary of ML prediction results for breaking wave depth 121
Table 5.1. RF prediction results for breaking wave height 133
Table 5.2. RF prediction results for breaking wave depth 139
Table 5.3. ANN prediction results for breaking wave height 147
Table 5.4. ANN prediction results for breaking wave height 149
Fig. 2.1. Illustrations of Machine Learning Algorithms 27
Fig. 2.2. Description of linear regression 29
Fig. 2.3. Concept of Support Vector Machine 35
Fig. 2.4. Structure of decision tree 36
Fig. 2.5. Random Forest schematic 38
Fig. 2.6. Flow chart of Bagging 39
Fig. 2.7. General layout of neural network 41
Fig. 2.8. Step function 43
Fig. 2.9. Sigmoid function 44
Fig. 2.10. tanh function 45
Fig. 2.11. ReLU function 46
Fig. 2.12. Leaky ReLU function 47
Fig. 3.1. Definition sketch of wave at incipient breaking 50
Fig. 3.2. Relationship between Hb/hb, Hb/Lo and Ho/Lo by historical laboratory experiments[이미지참조] 60
Fig. 3.3. Relationship between Hb/Lo, hb/Lo and Ho/Lo by historical laboratory experiments[이미지참조] 61
Fig. 3.4. Relationship between wave breaker index and deep water wave steepness by historical laboratory experiments 63
Fig. 3.5. Relationship between wave breaker index and deep water wave steepness by beach slope 64
Fig. 3.6. Comparison of wave breaking height formula LeMehaute and Koh(1967) with experimental data 66
Fig. 3.7. Comparison of wave breaking height formula Rattanapitikon and Shibayama(2006) with experimental data 66
Fig. 3.8. Comparison of wave breaking height formula Camenen and Larson(2007) with experimental data 67
Fig. 3.9. Comparison of wave breaking height formula Goda(2010) with experimental data 67
Fig. 3.10. Comparison of wave breaking depth formula Rattanapitikon and Shibayama(2006) with experimental data 72
Fig. 3.11. Comparison of wave breaking depth formula Xie et al.(2019) with experimental data 72
Fig. 4.1. k-fold cross-validation 75
Fig. 4.2. Grid and random search of nine trials for hyper-parameter tuning 76
Fig. 4.3. Process for cross-validation and hyperparameter optimization using Scikit-learn 77
Fig. 4.4. Predicted wave breaking height using LRM 81
Fig. 4.5. Predicted wave breaking depth using LRM 82
Fig. 4.6. Predicted wave breaking height using RR for α=0.01 85
Fig. 4.7. Predicted wave breaking height using RR for α=0.1 85
Fig. 4.8. Predicted wave breaking height using RR for α=0.2 86
Fig. 4.9. Predicted wave breaking height using RR for α=0.3 86
Fig. 4.10. Predicted wave breaking depth using RR for α=0.01 88
Fig. 4.11. Predicted wave breaking depth using RR for α=0.1 88
Fig. 4.12. Predicted wave breaking depth using RR for α=0.2 89
Fig. 4.13. Predicted wave breaking depth using RR for α=0.3 89
Fig. 4.14. Predicted wave breaking height using HR for α=0.5, є=1.0 92
Fig. 4.15. Predicted wave breaking height using HR for α=0.5, є=3.0 93
Fig. 4.16. Predicted wave breaking height using HR for α=0.5, є=5.0 93
Fig. 4.17. Predicted wave breaking height using HR for α=1.0, є=1.0 94
Fig. 4.18. Predicted wave breaking height using HR for α=1.0, є=3.0 94
Fig. 4.19. Predicted wave breaking height using HR for α=1.0, є=5.0 95
Fig. 4.20. Predicted wave breaking depth using HR for α=0.5, є=1.0 97
Fig. 4.21. Predicted wave breaking depth using HR for α=0.5, є=3.0 97
Fig. 4.22. Predicted wave breaking depth using HR for α=0.5, є=5.0 98
Fig. 4.23. Predicted wave breaking depth using HR for α=1.0, є=1.0 98
Fig. 4.24. Predicted wave breaking depth using HR for α=1.0, є=3.0 99
Fig. 4.25. Predicted wave breaking depth using HR for α=1.0, є=5.0 99
Fig. 4.26. Predicted wave breaking height using RANSAC for iter=50 102
Fig. 4.27. Predicted wave breaking height using RANSAC for iter=100 103
Fig. 4.28. Predicted wave breaking height using RANSAC for iter=150 103
Fig. 4.29. Predicted wave breaking height using RANSAC for iter=200 104
Fig. 4.30. Predicted wave breaking height using RANSAC for iter=300 104
Fig. 4.31. Predicted wave breaking height using RANSAC for iter=500 105
Fig. 4.32. Predicted wave breaking depth using RANSAC for iter=50 107
Fig. 4.33. Predicted wave breaking depth using RANSAC for iter=100 107
Fig. 4.34. Predicted wave breaking depth using RANSAC for iter=150 108
Fig. 4.35. Predicted wave breaking depth using RANSAC for iter=200 108
Fig. 4.36. Predicted wave breaking depth using RANSAC for iter=300 109
Fig. 4.37. Predicted wave breaking depth using RANSAC for iter=500 109
Fig. 4.38. Predicted wave breaking height using SVM for C = 1.0 112
Fig. 4.39. Predicted wave breaking height using SVM for C = 5.0 113
Fig. 4.40. Predicted wave breaking height using SVM for C = 10.0 113
Fig. 4.41. Predicted wave breaking height using SVM for C = 20.0 114
Fig. 4.42. Predicted wave breaking height using SVM for C = 50.0 114
Fig. 4.43. Predicted wave breaking height using SVM for C = 100.0 115
Fig. 4.44. Predicted wave breaking depth using SVM for C = 1.0 117
Fig. 4.45. Predicted wave breaking depth using SVM for C = 5.0 117
Fig. 4.46. Predicted wave breaking depth using SVM for C = 10.0 118
Fig. 4.47. Predicted wave breaking depth using SVM for C = 20.0 118
Fig. 4.48. Predicted wave breaking depth using SVM for C = 50.0 119
Fig. 4.49. Predicted wave breaking depth using SVM for C = 100.0 119
Fig. 4.50. Comparison of wave breaking height of all data sets using new formula 124
Fig. 4.51. Comparison of wave breaking height of all data sets using LeMahaute and Koh(1967) formula 124
Fig. 4.52. Comparison of wave breaking height of all data sets using Rattanipitikon and Shibayama(2006) formula 125
Fig. 4.53. Comparison of wave breaking height of all data sets using Carmenen and Larson(2007) formula 125
Fig. 4.54. Comparison of wave breaking height of all data sets using Goda(2010) formula 126
Fig. 4.55. Comparison of wave breaking depth of all data sets using new formula 128
Fig. 4.56. Comparison of wave breaking depth of all data sets using Rattanipitikon and Shibayama(2006) formula 128
Fig. 4.57. Comparison of wave breaking depth of all data sets using Xie et al.(2019) formula 129
Fig. 5.1. Predicted wave breaking height using RF for Dep=4, No=50 134
Fig. 5.2. Predicted wave breaking height using RF for Dep=4, No=100 134
Fig. 5.3. Predicted wave breaking height using RF for Dep=4, No=200 135
Fig. 5.4. Predicted wave breaking height using RF for Dep=6, No=50 135
Fig. 5.5. Predicted wave breaking height using RF for Dep=6, No=100 136
Fig. 5.6. Predicted wave breaking height using RF for Dep=6, No=200 136
Fig. 5.7. Predicted wave breaking height using RF for Dep=8, No=50 137
Fig. 5.8. Predicted wave breaking height using RF for Dep=8, No=100 137
Fig. 5.9. Predicted wave breaking height using RF for Dep=8, No=200 138
Fig. 5.10. Predicted wave breaking depth using RF for Dep=4, No=50 140
Fig. 5.11. Predicted wave breaking depth using RF for Dep=4, No=100 141
Fig. 5.12. Predicted wave breaking depth using RF for Dep=4, No=200 141
Fig. 5.13. Predicted wave breaking depth using RF for Dep=6, No=50 142
Fig. 5.14. Predicted wave breaking depth using RF for Dep=6, No=100 142
Fig. 5.15. Predicted wave breaking depth using RF for Dep=6, No=200 143
Fig. 5.16. Predicted wave breaking depth using RF for Dep=8, No=50 143
Fig. 5.17. Predicted wave breaking depth using RF for Dep=8, No=100 144
Fig. 5.18. Predicted wave breaking depth using RF for Dep=8, No=200 144
Fig. 5.19. Mean square error of training data as a function of the number of hidden layer neurons 146
Fig. 5.20. Comparison of the predicted wave breaking heights applying 1-hidden layer ANN results for input parameters 148
Fig. 5.21. Comparison of the predicted wave breaking heights applying 2-hidden layer ANN results for input parameters 149
Fig. 5.22. Comparison of the predicted wave breaking depth applying 1-hidden layer ANN results for input parameters 151
Fig. 5.23. Comparison of the predicted wave breaking depth applying 2-hidden layer ANN results for input parameters 151
Fig. 5.24. Wave breaking height of test data sets using Machine Learning 152
Fig. 5.25. Wave breaking height of test data sets using experiment formula 153
Fig. 5.26. Wave breaking depth of test data sets using Machine Learning 154
Fig. 5.27. Wave breaking depth of test data sets using experiment formula 154