실시간 수재해 예측과 대응을 위한 수치모형과 기계학습의 연계 = Linking Hydraulic Model with Machine Learning Approach for Real-time Flood Disaster Prediction and Response

다국어 초록 (Multilingual Abstract)

Sudden flooding has been on the rise in the watersheds that have undergone significant urbanization, and this is attributable to climate change. This trend presents the potential of tremendous property damage and losses of lives, and diverse methods have been applied to address the issue with greater efficiency. Unlike the traditional numerical analysis models and statistical analysis methods, flood analysis using machine learning is a data-based analysis that can present results in a short period of time through prior learning. Flood analysis based on machine learning involves the use of a wide range of simulation-based and observational data, and thus may be viewed as a type of data-based analysis. Accordingly, it is important to apply rational machine learning techniques along with data that are appropriate in terms of quality and quantity, and this technique may be actively applied to analyze and respond to unforeseen hydrological events and disasters. In this study, a machine learning technique with the use of one- and two-dimensional numerical analysis models, diverse hydrological observation data, and deep learning was applied.
As for the water disaster analysis and prediction, an analysis of the upper and lower streams of the Paldang dam and a flood analysis of a single urban drainage basin were carried out. With respect to an inflow of dams with a return period of 200 years or more, a flood map prediction for rising water levels of the Han-river was made with the application of the random-forest regression technique, and this is considered to be an efficient method to determine in real time the flooding pattern for the excessive dam inflow that can be expected. Using the range of flooding according to the amount of dam inflow under the probable maximum flood (PMF) condition, the utility of the flooding range prediction technique was proposed. The order of priority of the flood response to be made by each district of Seoul to counter the economic damage resulting from flooded buildings and casualties, and the level of importance of each factor affecting flood damage was estimated using the random forest technique. The proposed method can be used to select the districts in which flood response and restoration must be performed first in the event that a large-scale flood with a recurrence interval of 200 years or more occurs in the watershed of a metropolitan city.
The flood analysis of urban drainage basin units was conducted on Samseong drainage basin, and a LSTM (Long Short Term Memory) neural network model capable of predicting the total amount of overflow in real time according to the occurrence of rainfall events was trained. A comparison with the verified one-dimensional urban runoff analysis model showed that the LSTM neural network model could even accurately model the peak overflow. Using the predicted total overflow, the flooding range and flood risk map were predicted. When predicting the flooding range, the range was predicted based on the flooding depth using the logistic regression technique, and while it failed to produce a high area fitness for a relatively high flooding depth, it was deemed to be a highly appropriate method for predicting the overall flooding range according to the rainfall intensity. For the analysis of the flood risk map, the formula for calculating the flood risk level presented in the Department for Environment, Food & Rural Affairs, Environment Agency (DEFRA)’s Flood Risks to People report was used, and the flood risk level was determined in grid units using the flooding depth and flow rate information obtained from the two-dimensional flood analysis. A flood risk map was generated by taking into consideration both scenarios in which the debris factor associated with water disasters was and was not considered, and predictions were made through random-forest classifications. The results of the prediction and analysis showed that a relatively high flood risk level was found in residential areas with many alleys when the debris factor was taken into account. The prediction results using the random forest were compared with the results of using the flood risk calculation formula, and a high degree of agreement was found. Thus, it was judged that suggested methodology could be used as providing the useful data for performing flood control of each urban drainage basin at a later time.
Through flood prediction and analysis using diverse machine learning, it is possible to determine the danger zones in real-time in the event of a water disaster. The proposed methodology can be said to have high reliability and accuracy, as it uses a database built based on observational data and the results of applying verified numerical analysis models. Its practicality was also demonstrated by presenting the method of utilization with respect to the prediction and analysis results, and it is expected that the prediction model will help speed up the processes of the urban watershed flood control system if it is improved with more basic data. The aim of this study was to demonstrate the diverse uses of various observational data and numerical analysis models in connection with machine learning techniques and that they are accurate enough to be actively applied in practice to defend against flooding.

목차 (Table of Contents)

• 제 1 장 서 론 1
• 1.1 연구배경 1
• 1.2 연구목적 4
• 1.3 연구동향 6
• 1.4 연구내용 10
• 제 2 장 수치해석과 기계학습의 연계 홍수분석 13
• 2.1 수치해석 모형 13
• 2.1.1 DAMBRK 13
• 2.1.2 SWMM 15
• 2.1.3 FLO-2D 19
• 2.2 기계학습 모형 23
• 2.2.1 로지스틱 회귀 24
• 2.2.2 랜덤 포레스트 26
• 2.2.3 LSTM(Long Short-Term Memory) 신경망 30
• 제 3 장 연구 대상유역에 대한 입력자료 구축 34
• 3.1 연구 대상유역 35
• 3.1.1 홍수범람 분석 35
• 3.1.2 내수침수 분석 40
• 3.2 하천에서의 외수범람해석 43
• 3.2.1 DAMBRK 모형의 구축 43
• 3.2.2 홍수범람해석 결과 46
• 3.3 도시지역에서의 내수침수해석 56
• 3.3.1 SWMM의 구축 및 도시유출해석 56
• 3.3.2 침수해석결과 홍수위험등급지도 구축 67
• 3.4 기계학습을 위한 데이터베이스의 구축 75
• 제 4 장 수재해 분석 및 예측 77
• 4.1 홍수위 영향에 따른 홍수범람도 예측 77
• 4.1.1 계획빈도 유입량에 따른 최고홍수위 추정기법 77
• 4.1.2 임의 빈도에 대한 침수범위 추정 95
• 4.2 실시간 내수침수 및 위험등급 예측 107
• 4.2.1 LSTM 모형 기반 월류량 예측 107
• 4.2.2 로지스틱 회귀와 연계한 범람범위 예측 118
• 제 5 장 비교검토 및 활용방안 125
• 5.1 인명 및 건물 피해 기준 위험 행정구역 선정 129
• 5.2 랜덤 포레스트 기반 홍수위험등급 예측 149
• 제 6 장 결 론 168
• 참 고 문 헌 173
• Abstract 181

참고문헌 (Reference)

1. , 기상청, 2018년 이상기후 보고서, , 2019

2. 치수관리 정책, 서울시, 서울 연구원, , 2015

3. “수치모형을 이용한 순차적 댐 붕괴 모의.”, 최현구, 한건연, 박세진, 대한토 목학회논문집, 제33권, 제2호, pp. 1797-1807, , 2013

4. “Sub-Grid 기법을 도입한 2차원 도시침수 해석.”, 한건연, 이경택, 안기홍, 조완희, 대한토목학회 학술대회, pp. 1329-1332, , 2007

5. “딥러닝 오픈 라이브러리를 이용한 하천수위 예측.”, 이경상, 정성호, 이대업, 한국방재학회논문집, 제18권, 제1호, pp. 1-11, , 2018

6. “신경망을 이용한 낙동강 유역 홍수기 댐유입량 예 측.”, 신현석, 윤강훈, 서봉철, 한국수자원학회논문집, 제37권, 제1호, pp. 67-75, , 2004

7. “머신러닝을 이용한 공공시 설 호우피해 예측함수 개발.”, 김형수, 이명진, 김종성, 최창현, 박희경, 박기혁, 한국방재학회논문집, 제17권, 제6호, pp. 443-450, , 2017

8. “도시침수 해석을 위한 동적 인공신경망의 적용 및 비교.”, 금호준, 김현일, 한건연, 대한토목학회논문집, 제38권, 제5호, pp. 671~683, , 2018

9. “로지스틱 회귀에 의한 도시 침수발생의 한계강우량 산정.”, 한건연, 김현일, 대한토목학회논문집, 제39권, 제6호, pp. 713-723, , 2019

10. “팔당댐 방류량에 따른 한강 시민공원의 수리학적 영향 분석.”, 곽창재, 이상원, 이재준, 한국방재학회논문집, 제8권, 제6호, pp. 101~111, , 2008

11. “홍수범람지도작성을 위한 GIS 기반 1차원 동역학 적 해석기법.”, 김병현, 한건연, 최승용, 한국방재학회논문집, 제11권, 6호, pp. 227~235, , 2011

12. “유역 모형과 하천 모형의 연계를 통한 낙동강 본류 흐름 예측.”, 강형식, 한 국수자원학회논문집, 제33권, 제2호, pp. 577~590, , 2011

13. “도시지역 도로 네트워크를 활용한 침수지역 예측 에 관한 연구.”, 한건연, 손아롱, 김병현, 대한토목학회논문집, 제35권, 제2호, pp. 307-318, , 2015

14. “피해이력 및 유역특성을 고려한 도시침수 위험기 준 설정 및 적용, 강호선, 배창연, 조재웅, ” 한국수자원학회논문집, 제51권, 제1호, pp. 1-10, , 2017

15. “수리학적 기법과 확률론적 방법을 연계한 실시간 도시침수 예측기 법.”, 김현일, 석사학위 논문, 경북대학교, , 2017

16. 저지대 홍수피해 저감을 위한 도시계획기법 연구 – 서울시 사례를 중심으로, 고태규, 이원영, 서울도시연구, Vol. 13, No. 4, pp. 287~300, , 2012

17. “실시간 자료지향형 예측과 2차원 수리해석을 연계한 도시침수 저 감모형 개발.”, 손아롱, 박사학위 논문, 경북대학교, , 2013

18. “팔당댐 방류량과 황해(서해) 조석영향에 따른 팔당댐 하류 부 수위상승도달시간 예측.”, 이재홍, 이정규, 한국방재학회논문집, 제10권, 제2호, pp. 111-122, , 2010