표제지
요약
목차
1. 서론 15
1.1. 연구배경 및 목적 15
1.2. 연구내용 및 구성 18
2. 이론적 고찰 20
2.1. 시단위 데이터의 활용 20
2.2. Attention mechanism 23
2.3. 결측치 대체 방법론 29
2.3.1. Multi-directional recurrent neural network(M-RNN) 29
2.3.2. Bidirectional recurrent imputation for time series(BRITS) 33
2.3.3. Self-attention-based imputation for time series(SAITS) 39
2.4. 다지점 시공간 예측 모델 46
2.4.1. Node embedding 46
2.4.2. Spatio-temporal graph attention network(ST-GRAT) 51
3. 연구 방법 56
3.1. 연구 대상지 및 자료 수집 59
3.2. 결측치 대체 모델 구축 62
3.2.1. 시계열 특성을 고려한 결측치 대체 모델 개발 63
3.2.2. 결측치 대체 비교모델 구축 65
3.3. 다지점 수질 예측을 위한 시공간 모델 개발 67
3.3.1. 노드 임베딩(Node embedding) 67
3.3.2. 다지점 수질 예측 모델 구축 69
4. 연구 결과 및 고찰 72
4.1. 시단위 수질 데이터 특성 및 현황 72
4.2. 모델링 결과 75
4.2.1. 결측치 대체 모델 성능 평가 75
4.2.2. 시공간 분석을 통한 다지점 수질 예측 82
4.2.3. 지점 간 어텐션 기여도 해석 101
4.3. 고찰 106
4.3.1. 센서 데이터의 활용 가능성 106
4.3.2. 상하류를 포함한 다지점 수질 예측 108
4.3.3. 모델별 시공간 어텐션 기여도 110
5. 결론 112
References 114
Appendix 122
ABSTRACT 126
Table 2.1. Summary of popular score functions 24
Table 2.2. Four criteria for categorizing attention mechanism 26
Table 3.1. Missing status of automatic measurement station data 62
Table 3.2. Hyperparameter for SAITS model 64
Table 3.3. Hyperparameter for M-RNN model 65
Table 3.4. Hyperparameter for BRITS model 65
Table 3.5. T/M measurement stations matched with Automated measurement stations 68
Table 3.6. Hyperparameter tuning range for ST-GRAT 70
Table 4.1. Performance of missing value imputation models under various artificial missing rates. 75
Table 4.2. Performance evaluation of multi-site TOC concentration prediction using ST-GRAT. 84
Table 4.3. Performance evaluation of multi-site Chlorophyll a concentration prediction using ST-GRAT. 85
Figure 1.1. Flowchart of research process 19
Figure 2.1. Status of Automatic Measurement Stations (NIER) 21
Figure 2.2. The general scheme of attention model 23
Figure 2.3. Comparison of attention and self-attention mechanism 27
Figure 2.4. Structure of multi-head attention mechanism 28
Figure 2.5. Structure of M-RNN 30
Figure 2.6. Description of DMSA 39
Figure 2.7. Structure of SAITS 42
Figure 2.8. Exampe of first order and second order proximity 46
Figure 2.9. Sampling neighborhood 48
Figure 2.10. Aggregate feature information from neighbors 49
Figure 2.11. Spatial and temporal attention in graph networks 53
Figure 2.12. Structure of ST-GRAT 54
Figure 3.1. Schematic diagram of the research process 58
Figure 3.2. Map of sampling sites in Nakdong river 60
Figure 3.3. Process of making weighted adjacency matrix 69
Figure 4.1. Variation of TOC concentration 73
Figure 4.2. Variation of Chlorophyll a concentration 73
Figure 4.3. Trend of missing values 74
Figure 4.4. Time series plots of TOC missing value imputation using SAITS at 20% artificial missing rate. Observed values are represented by... 79
Figure 4.5. Time series plots of Chlorophyll a missing value imputation using SAITS at 20% artificial missing rate. Observed values are... 81
Figure 4.6. Time series plots of TOC concentration prediction 12 hours ahead using ST-GRAT. Blue dots represent observed values, while red... 87
Figure 4.7. Time series plots of TOC concentration prediction 48 hours ahead using ST-GRAT. Blue dots represent observed values, while red... 88
Figure 4.8. Time series plots of Chlorophyll a concentration prediction 12 hours ahead using ST-GRAT. Blue dots represent observed values, while... 89
Figure 4.9. Time series plots of Chlorophyll a concentration prediction 48 hours ahead using ST-GRAT. Blue dots represent observed values, while... 90
Figure 4.10. Time series plots of TOC concentration prediction 12 hours ahead using LINE+ST-GRAT. Blue dots represent observed values, while... 92
Figure 4.11. Time series plots of TOC concentration prediction 48 hours ahead using LINE+ST-GRAT. Blue dots represent observed values, while... 93
Figure 4.12. Time series plots of Chlorophyll a concentration prediction 12 hours ahead using LINE+ST-GRAT. Blue dots represent observed values,... 94
Figure 4.13. Time series plots of Chlorophyll a concentration prediction 48 hours ahead using LINE+ST-GRAT. Blue dots represent observed... 95
Figure 4.14. Time series plots of TOC concentration prediction 12 hours ahead using GraphSAGE+ST-GRAT. Blue dots represent observed values,... 97
Figure 4.15. Time series plots of TOC concentration prediction 48 hours ahead using GraphSAGE+ST-GRAT. Blue dots represent observed values,... 98
Figure 4.16. Time series plots of Chlorophyll a concentration prediction 12 hours ahead using GraphSAGE+ST-GRAT. Blue dots represent... 99
Figure 4.17. Time series plots of Chlorophyll a concentration prediction 48 hours ahead using GraphSAGE+ST-GRAT. Blue dots represent... 100
Figure 4.18. Encoder-decoder temporal attention at the peak concentration point in Juckpo. TOC value was computed on June 24, 2022, at 00:00 hours, and Chlorophyll a value was computed on August 9, 2022, at 20:00 hours. 102
Figure 4.19. The coefficient of variation for the encoder-decoder temporal attention of each model in predicting TOC and Chlorophyll a. 103
Figure 4.20. Decoder spatial attention at the peak concentration point in Juckpo. TOC value was computed on June 24, 2022, at 00:00 hours, and Chlorophyll a value was computed on August 9, 2022, at 20:00 hours. 105