미세먼지 예측을 위한 딥러닝 기반 농도별 분리 예측 모델

조경우 한국기술교육대학교 일반대학원 2020 국내박사

본 논문에서는 기계 학습을 사용한 미세먼지 예측 모델의 저농도 미세 먼지 과대 예측 현상과 고농도 미세먼지 예보정확도 향상을 위해 미세먼지 저농도, 고농도를 분류하는 분류 모델과 미세먼지 농도를 예측하는 저농도, 고농도 예측 모델의 세 모델을 결합하여 예측을 수행하는 딥러닝 기반 농도별 분리 예측 모델을 제안하였다. 이를 위해 천안 지역의 10년간 시간별 기상 및 대기오염 물질 데이터를 수집하여 예측 모델의 입력 데이 터로 활용할 수 있도록 데이터를 분류 및 전처리하였다. 제안 모델에 사용된 딥러닝 알고리즘은 심층 신경망(Deep Neural Network)을 기반으로 하였으며, 미세먼지 농도 분류 모델의 경우 고농도 경계인 81[  ]을 기준 으로 저농도와 고농도의 이진 분류를 수행한다. 이후 분류 모델의 결과에 따라 농도별 예측 모델에 해당 입력변수를 전달하여 최종 미세먼지 농도를 예측한다. 각 모델의 하이퍼 파라미터의 경우 모델별로 4,800개의 조합을 가진 그리드 탐색을 수행하였으며, 모델의 일반화 성능을 위해 5개의 fold로 training data를 나누어 모델별로 최적 하이퍼 파라미터를 선정하였 다. 모델의 입력변수의 경우, 기상 및 대기오염 물질과 미세먼지 농도와의 농도별, 계절별 상관분석을 통해 역할별 모델의 입력변수를 선정하였으며, 고농도 미세먼지와 상관성을 보인 변수 사용의 효과를 확인하기 위해 전체 모델에 사용된 모든 변수를 사용한 고농도 예측 모델과 선정 입력변수를 사용한 모델 간의 예측 성능 비교를 수행하였다. 성능 비교 결과 전체 입력변수를 사용한 모델의 RMSE는 15.9537로 선정 입력변수를 사용한 모델의 16.1014보다 더 좋은 성능을 나타내었으나, 103회의 저농도 예측이발생되어 미세먼지 농도를 상대적으로 과소예측하는 경향을 보였다. 선정 입력변수를 사용한 모델의 경우, 0회의 저농도 예측 횟수를 나타내었으며 AQI ‘매우 나쁨’ 정확도의 경우 전체 입력변수 모델에 비해 18.86[%] 향상된 정확도를 나타내어 고농도 미세먼지와 상관성이 있는 변수 사용을통해 예측 성능을 향상시킬 수 있었다. 제안 모델의 예측 성능 분석을 위해 단일 신경망 모델인 DNN, RNN, LSTM 모델을 설계하고 예측 성능을 비교하였다. 예측 성능 비교 결과 단일 신경망 모델들에 비해 제안 모델의 RMSE, MAPE, 상관 값의 성능 이 근소하게 낮았다. 그러나, 단일 신경망 모델들의 경우 AQI ‘보통’ 정확 도에 비해 타 AQI 구간의 정확도가 낮은 모습을 보였으며, 제안 모델의 경우 AQI 전 영역에 대해 80[%] 이상의 고른 정확도를 가졌다. 특히, 고농도 기준인 81[  ] 이상 고농도 미세먼지 예측 정확도의 경우 단일신경망 모델의 최고 성능보다 5.73[%] 향상된 81.87[%]를 나타내었다. 또한, 미세먼지 데이터의 농도별 불균형을 고려하여 g-mean을 통한 예측성능 비교 시, 0.87∼0.89의 값을 나타낸 단일 신경망 모델에 비해 제안 모델의 경우 0.9133로 높았다. 따라서, 제안하는 농도별 분리 예측 모델은 단일 신경망 모델들의 저농도 과대 예측 문제를 완화할 수 있으며, 전체 미세먼지 데이터에서 낮은 비율을 차지하는 고농도 미세먼지 예측 문제에효과적임을 확인하였다. To mitigate the overprediction of low-concentration particulate matter and enhance the prediction accuracy of high-concentration particulate matter in particulate matter prediction models using machine learning, we herein proposed a concentration-classified prediction model. This model is based on deep learning and combines (i) a classification model that classifies low-concentration and high-concentration particulate matter and (ii) low-concentration and high-concentration prediction models that predict particulate matter concentration. For this purpose, hourly weather and air pollutant data over 10 years in the Cheonan region were collected, and the data were classified and preprocessed for use as input data of the prediction model. The deep learning algorithm used in our proposed model was based on a deep neural network, and the classification model for particulate matter concentration conducts binary classification of low-and high-concentration particulate matter based on a high-concentration boundary of 81[  ]. Subsequently, according to the results of the classification model, the corresponding input variable is transmitted to the concentration-classified prediction models and the final particulate matter concentration is predicted. In terms of the model hyperparameters, grid search was conducted for each model with 4,800 combinations, and for the generalization performance of the models, the training data were divided into five folds and optimal hyperparameters were selected for each model. The input variables of the models were selected by role through a correlation analysis according to the season and concentration between weather and air pollutants and particulate matter concentration. To confirm the effect of using variables that correlated with high-concentration particulate matter, in this study, we compared the prediction performance between the high-concentration prediction model using all variables applied in all models and the model using the selected input variables. The performance comparison showed that the RMSE of the model using all input variables was 15.9537, exceeding the performance of the model using the selected input variables of 16.1014. However, a total of 103 low-concentration predictions occurred, indicating that it tended to underestimate the particulate matter concentration. The model using the selected input variables showed zero low-concentration predictions and improved accuracy of 18.86[%] compared to the model using all input variables in terms of AQI “very bad” accuracy. These results confirmed that the prediction performance can be improved by using variables correlating with high-concentration particulate matter. To analyze the prediction performance of the proposed model, DNN, RNN, and LSTM models, which are single-neural-network models, were designed and their prediction performances were compared. According to the comparison of prediction performance, the proposed model exhibited slightly lower performance than the single-neural-network models for RMSE, MAPE, and correlation value. However, the single-neural-network models showed lower accuracy for other AQI sections than AQI “normal” accuracy, whereas the proposed model showed accuracy of at least 80[%] for all AQI sections. Particularly, in terms of high-concentration particulate matter prediction accuracy above the high-concentration standard of 81 [  ], the proposed model exceeded the maximum performance of the single-neural-network models by 5.73[%] at 81.87[%]. Furthermore, comparing the prediction performance through g-mean considering the imbalance of particulate matter data by concentration, the proposed model showed a high value of 0.9133 compared to the single-neural-network models at 0.87–0.89. Thus, our proposed concentration-classified prediction model can mitigate the overprediction of low-concentration particulate matter in single-neural-network models, and it was confirmed to be effective for predicting high-concentration particulate matter.

건물의 소비전력 절감을 위한 위치 기반 전원제어 시스템 연구

조경우 한국기술교육대학교 대학원 2015 국내석사

The study proposed a system that controls the power for electronic devices based on the user location by using fingerprinting technique, an indoor localization system, to cut down power consumption of a building. To this end, it developed a data collection application for a training stage of the finger-printing technique, selected the cell intervals of 80[cm], 160[cm], 300[cm], the multiples of 20[cm] that have the least variable width and collects periodic RSSI in the environment when reference map is built and established a reference map by using circular intersecting points from each AP. Regarding RSSI that goes beyond ±6[dBm] from the accumulated RSSI average collected in the actual localization stage, proposed localization algorithm based on accumulated average RSSI that receives a new RSSI vector through ACK signal and designed the localization system. It also monitored power controlling of indoor lights to realize the power control like home plug and developed the program and power control module for control, which has control over a total of eight power sources. It was found from the experiment after creating the environment for the lab and corridor that adjacent coordinates that located the user were accurately computed compared with the existing localization technique and the coordinates were used to control power. A portable integrating wattmeter was used for another experiment in the environment that has four people. The result showed that the existing light recorded an integrated power of about 145[Wh] while it was about 25[Wh] when four were located in the same lighting area and 83[Wh] when they are in the different lighting area with the proposed method. In this light, it was identified to be more effective to reduce power consumption than the existing method. 본 논문에서는 건물의 소비전력 절감을 위해 실내 위치 측위 기법인 fingerprinting 기법을 사용하여 사용자의 위치를 기반으로 가전기기의 전원을 제어하는 전원제어 시스템을 제안하였다. 이를 위해 fingerprinting 기법의 트레이닝 단계를 위한 데이터 수집 어플리케이션을 개발하고, reference map 구축 시 실제 환경에서 변화폭이 가장 낮고, 주기적인 RSSI가 수집되는 20[cm]의 배수를 택하여 80[cm], 160[cm], 300[cm]의셀 간격을 선택, 각 AP로부터 원형의 교차점을 이용하여 reference map 을 구축하였다. 실제 측위단계에서 수집되는 누적 RSSI 평균에 ±[dBm] 을 벗어난 RSSI를 에러로 간주, ACK 신호를 통해 새로운 RSSI 벡터를 수신 받는 누적평균 RSSI 기반 위치추정 알고리즘을 제안하고 위치측위 시스템을 설계하였다. 또한, 홈 플러그와 같은 전원제어를 구현하기 위해 실내조명을 대상으로 전원제어 상황을 모니터링 하고, 제어하기 위한 프로그램 및 전원제어 모듈을 개발하였다. 개발된 전원제어 모듈은 총 8개의 전원 소스 제어가 가능하다. 연구실과 복도를 대상으로 실험환경을 구축한 후 실험한 결과, 사용자가 존재하는 인접 좌표를 기존의 위치 추정 방식보다 정확하게 추정하는 것을 확인하였으며, 해당 위치 좌표를 이용 하여 전원을 제어하였다. 휴대용 전력적산계를 사용하여 4인이 존재하는 환경에서 실험한 결과, 기존 조명의 경우 약 145[Wh]의 적산전력을 나타낸 반면, 제안한 방식을 사용했을 때 4인이 동일한 조명 영역에 위치한 경우 약 25[Wh], 서로 다른 조명 영역에 위치한 경우 약 83[Wh]로 기존 방식에 비해 소비전력 절감 효과가 나타나는 것을 확인하였다.

海運企業의 電子商去來 活用 및 成果에 關한 實證硏究 : 定期船海運營業을 中心으로

조경우 韓國海洋大學校 2006 국내박사

After the introduction of Internet, cyber space has played a new role as a competitive advantage for shipping companies. For traditional shipping industry, extended Internet meant to prosper in E-commerce. For example, Schedule confirmation, booking, bill of lading, cargo trace, and arrival notice etc. are the major cyber logistics services which make the computers work. Particularly Korean shipping companies had to figure out how to be competent internationally through introducing and activating E-commerce as soon as possible. Under this circumstance, this study searched main reasons that affect the embodiment for Korean shipping companies' E-commerce and could analyze the application and its standard of E-commerce positively. Based on the result drew upon, the purpose of this study was to suggest logical foundation and points which were necessary for establishing E-commerce of Korean shipping companies' development strategy and activation plan. So as to do this, firstly, general inquiries in the document study were made into E-commerce for its outline with E-commerce concept, pattern, present condition and prospects for future. Moreover, deep study for proceeded researches should arrange and evaluate E-commerce related model studies and corroborations as well as an application of EDI and Internet by transaction cost theory, innovation diffusion theory, strategic management theory, and inter-organizational system, etc. so that this study could deduce its study models and useful points for the hypothesis. Secondly, in empirical study, study models and hypothesis for empirical analysis were established based on logical foundations from the document study. The composition of questions, selection of samples, data collection and collected material analysis were confirmed and carried out. Established research model and hypothesis were utilized in the previous studies or Korean shipping companies were examined using questions developed from the concept mentioned in similar studies. Then this study verified through analysis of 148 collected responses and used computer statistic packages, SAS(The SAS System for Windows v9.0) and LISREL(LISREL for Window v8.12a). For empirical study, general characters of the investigated objects were looked into first and adequacy and trust of each variable and measurement tool shall be analysed for their main causes. Cronbach's Alpha(α) coefficient was used for the analysis. In order to do empirical analysis, two step analyses were adopted. Firstly, this study examines the effects of shipping company character, shipping transaction character, and shipping environment character on the usage of E-commerce in shipping companies including service business and real transaction. Secondly, this study examines the effects of usage of E-commerce on the performance in shipping companies including direct performance, indirect performance, and strategic performance. In the empirical analysis, research hypotheses are being tested mainly by a structural equation model(SEM). The results from a structural equation model(SEM), developed using LISREL(LISREL for Window v8.12a), provide a strong support for the hypothesized relations. Parametric method like not only structural equation model(SEM) but also multiple regression analysis was used and the following would be summarized as follows: (1) Both shipping company character and shipping transaction character didn't have positive effect on the service business usage's level of E-commerce in shipping companies, but the service business usage's level of E-commerce in shipping companies was found to link with shipping environment character. That was, higher government's promotion policy and exterior support, greater the service business usage's level of E-commerce. (2) Also, both shipping company character and shipping transaction character had positive effect on the real transaction usage's level of E-commerce in shipping companies, but the real transaction usage's level of E-commerce in shipping companies was not found to link with shipping environment character. (3) Futhermore, E-commerce's usage performance in shipping companies including direct performance, indirect performance, and strategic performance was found to link with service business usage's level of E-commerce. But E-commerce's usage performance in shipping companies didn't have positive effect on the real transaction usage's level of E-commerce in shipping companies. That was, higher the service business usage's level of E-commerce, greater the E-commerce's usage performance composed of direct outcome, indirect outcome, and strategic outcome.

우리나라 市中銀行 小賣金融市場의 效率的 細分化 및 마케팅 戰略에 관한 硏究 : 先進 海外銀行의 小賣金融市場 및 마케팅 戰略과의 比較

조경우 漢陽大學校 經營大學院 2002 국내석사

After Korean economic crisis, some domestic banks have been troubled with the same bitterish failures and improper overseas investments, which overseas banks in USA and other countries experienced before, including damages from their accumulated dishonored bonds and currently, survive the crisis through the public funds and investments from overseas financial institutes in the course of domestic financial restructuring. In special, some banks, which were undertaken with wholesale banking affairs during huge M&A among banks, owned piles of dishonored bonds, consequently, resulting in nearly bankruptcy conditions. Still several poor banks are discussing additional M&A agreements with overseas financial institutes and domestic top-ranking banks. Also, the birth of mammoth banks from M&A deals among banks and the emergence of advanced overseas large financial institutes stimulate the heavy competition in the domestic financial market. In the meantime, domestic banks pursue scientific and systematic management systems to maintain VIP customers and to increase profit, variable sales environment and new marketing strategies for the security of good customers and benefits rising. And, to respond rapidly changing financial environment, painstaking management innovation and customer-oriented & state-of-the-art marketing concepts are required necessarily. For the purpose of the above, banks are reinforcing IR promotion like the system reengineering in the retail financial market and utilize various marketing materials like price-differentiation with all their effort. Meanwhile, the birth of mammoth banks during the domestic financial restructuring makes the domestic financial market more competitive and especially, some advanced overseas banks like CITI BANK and HSBC are looking for a chance to enter the domestic financial market with some aggressive marketing strategies. So, the domestic retail financial market is heavily competitive as the Age of the Warring States in Chinese History. From the account, the study reviewed the concept and status about the segmentation of the domestic retail banking market, bank's marketing service and Relationship Marketing and Private Banking theoretically and then, researched the system and conditions of the domestic retail financial market and the present status and strategies of overseas banks' retail banking market centering on cases. In addition, keeping pace with a sudden change in the financial circumstances, the study purposed to provide some courses in the future through the relative cases in order that domestic retail-banking market may secure the superiority and to represent the directions to create various strategies. IMF 이후 국내 일부 은행들은 과거 미국등 해외 은행들이 경험했던 도매 금융시장에서의 실패와 부적절한 해외투자로 인해 부도 채권이 누적되어 경영에 심각한 타격을 입었고 또한 국내 금융구조조정과정에서 정부의 공적자금 및 해외 금융기관의 투자지원을 통해 생존해 가고 있는 것이 일부 은행들의 현실이다. 특히 은행간 합병이 이루어지면서 도매 금융을 담당했던 일부은행들은 부실 채권이 누적되어 빈사상태에서 일부 은행은 타 금융기관에 합병 당하였으며 현재도 국내 부실 은행들은 국내 우량은행과 외국계 금융기관등에 M&A협의를 계속 진행하고 있다. 또한 은행간 인수 합병으로 국내 거대 은행의 탄생과 선진 해외 은행들의 국내시장의 적극 진출로 금융기관간의 경쟁을 더욱 치열하게 만들어가고 있으며 따라서 국내 은행들은 소매금융시장에서 우량고객 확보와 수익성 제고를 위한 과학적이고 체계적인 시스템 및 영업환경과 새로운 마케팅 전략을 구사하고 있다. 그리고 급속히 변화하는 금융환경에 대응하기 위해서는 뼈를 깎는 경영혁신과 고객 중심의 현대적 마케팅 개념이 그 어느 때 보다도 절실히 요구되고 있다. 이를 위해 은행들은 소매금융시장에서의 시스템 체제 정비 등 IR홍보에 주력하고 가격 차별화등 다양한 마케팅 수단을 총 동원하고 있다. 한편 금융 구조조정으로 인한 대형은행의 탄생으로 우리나라 소매금융시장의 경쟁이 더욱 치열해 지고 있으며 특히 CITI BANK, HSBC등 선진 해외 은행들이 최첨단 마케팅 기법으로 공격적인 국내시장 침투를 노리고 있으며 따라서 국내 소매금융시장의 국내‧해외 은행간의 춘추전국시대의 전쟁을 방불케 하고 있다. 따라서 본 연구는 국내 소매금융시장 세분화에 대한 개념과 현황 그리고 은행 마케팅 서비스 및 관계마케팅과 프라이빗 뱅킹에 관한 이론적 고찰을 살펴본 후 국내 소매금융시장의 제도 및 실태와 또한 해외 은행의 소매금융시장 현황과 전략에 대해 사례를 중심으로 연구하였다. 또한 급변하는 금융환경에 발 맞추어 국내은행 소매금융시장이 경쟁적 우위를 확보할 수 있도록 국내외 은행의 사례를 통해 향후 국내은행의 진로를 모색하고 여러 가지 전략을 강구하기 위한 방향을 제시하는데 그 목적이 있다 하겠다.

銀行서비스의 顧客 關係마케팅에 관한 硏究 : 地方金融市場을 中心으로

조경우 昌原大學校 經營大學院 2004 국내석사

In this study, a survey for banking consumers in Kyungsangnamdo was performed to analyze correlation between explanative variables and mediating variables of relationship marketing. Major results of the survey are as follows: First, consumers in Kyungsangnamdo keep neutrality for relationship discontinuation cost which is expectative loss to be incurred when they change transaction bank. Second, public values between the local citizen and their transaction banks bring about positive effect. Third, a communication tool, such as media advertisement, business letter, poster, or a staff's visit, seldom affects consumers choice of financial wares. Fourth, the great relationship commitment to transaction bank was not found. Fifth, the consumers in Kyungsangnamdo trust their transaction bank in staffs' ability, getting information, financial wares, computer ability, foreign exchange dealing ability, etc. In this study, cross tables are investigated to find the relative factors about the confidence or relationship commitment to transaction bank. Major results of the cross analysis are as follows: First, correlation between relationship discontinuation cost and relationship commitment was not found in Kyungsangnamdo. Second, correlation between contribution and relationship commitment was clearly found. This means that a bank has the more contribution to local society, consumers in Kyungsangnamdo have the more relationship commitment. Similarly, public value is distinctly related to consumers relationship commitment. Accordingly, financial consumers in Kyungsangnamdo have the strong relationship commitment which incurred from public value while they have the weak opportunity cost which incurred from changing transaction bank. Third, the more contribution of a bank is given to local society, the more trust of local consumers is given to the bank. Similarly, public values give trust to consumers. Fourth, correlation between communication and trust is found.

경기도 시각장애인의 생활체육 실태 조사 연구

조경우 용인대학교 2006 국내석사

본 연구의 목적은 경기도 시각장애인 생활체육 참여 실태조사를 통하여 체육 활동에 대한 인식과 참여 여부, 생활체육 형태, 생활체육활동 요인과 생활체육 환경, 그리고 생활체육 발전에 대한 시각장애인의 욕구와 여론 등을 조사․분석하는데 주목적이 있었다. 이러한 목적을 달성하기 위하여 경기도에 거주하는 시각장애인 110명을 편의표집하여 대상자로 선정하였고 1:1 면접을 통해 생활체육실태를 조사하였다. 시각장애인 생활체육실태를 분석하기 위한 본 설문지 내용은 첫째, 장애인 체육활동에 대한 인식과 참여태도영역, 둘째, 생활체육의 참여실태영역, 셋째, 생활체육 발전 방안 영역으로 구성하였다. 본 연구의 결론은 다음과 같다. 첫째, 경기도 시각장애인의 생활체육 참여 욕구는 높은 수준이었으며, 생활체육의 가치는 재활에서 건강과 체력증진으로 전환되었다. 둘째, 생활체육 편의시설의 확충과 보수, 프로그램 개발 및 보급, 전문지도자 양성 및 배치 등은 시각장애인의 생활체육 참여 활성화를 위한 중요한 필요조건이었다. 셋째, 정부의 행․재정적 지원과 일반 사람들과 차별 없는 행정을 요구하고 있었다. This study investigated the actual status of lifetime sports for people with visual impairment for sample of 110 from stratified cluster random sampling in Gyeonggi Province. The contents and scopes for the study are as followings. First, perceptions and attitudes toward physical activity and participation rate 1) physical fitness and health 2) Leisure time 3) participation rate. Second, condition and current status of lifetime sports for participation 1) for no participation 2) for participation. Third, needs and directions for future lifetime sports developments. The conclusion of this study is as follows; First, the desire for lifetime sports participation of people with visual impairment in lifetime sports was high level, the value of lifetime sports are known to change health and physical fitness promotion from rehabilitation. Second, the expansion and impairment sports facility of living athletics, program development and supply, training and disposition of professional leader etc. go known very important foundation in order to revitalize participation in lifetime sports for people with visual impairment. Third, it is going to present Governmental administrative and financial support without discrimination were stonly required.

한국과 중국간 트럭복합일관 수송체계(RFS)활성화 방안

조경우 인하대학교 국제통상물류대학원 2008 국내석사

Amiloride 및 몇가지 Aormone이 자라 膀胱의 短絡電流 및 電位差에 미치는 影響

조경우 전남대학교 1974 국내박사

Improving Streamflow Prediction with WRF-Hydro Modeling System using a Machine-learning Algorithm

조경우 연세대학교 일반대학원 2020 국내석사

물리 기반의 분포형 수문지면 모형을 활용한 하천 유출 예측을 위해서는 긴 시간 축적된 다양하고 정확한 입력데이터를 필요로 한다. 최근 개발된 WRF-Hydro 모형 또한 다양한 지역을 대상으로 활용되기 시작하였으나 주로 단기간 모의에 사용되어 왔다. 본 연구에서는 WRF-Hydro 모형의 하천 유출 예측 성능을 향상시키기 위해 머신러닝 알고리즘 적용하였으며, 특히 장기간 유출 예측 성능 향상시키고자 하였다. 우선 한강 상류에 위치한 소양댐의 유입량을 대상으로 WRF-Hydro 모형을 구축하였으며 PEST를 이용하여 보정된 매개변수를 사용하여 유입량을 모의하였다. 이후 LSTM이 사용된 머신러닝 알고리즘을 이용하여 WRF-Hydro 모형 모의값과 관측값의 잔차를 모형에 학습시켰으며, 학습된 머신러닝 모형은 WRF-Hydro모형의 하천 유출 예측 개선에 사용되었다. 2018년의 유입량을 3시간 단위로 모의한 결과를 소양댐 유입량자료와 비교하여 검증하였다. PEST로 최적화된 매개변수는 10일의 단기간 모의 결과를 향상시켰지만 1년의 장기간 모의 결과 개선에는 한계를 보였다. 그러나 이후 머신러닝 알고리즘이 적용된 WRF-Hydro 모형은 단기간 과 1년의 장기간 예측에서 모두 결과가 크게 개선되었다. 머신러닝 적용 전의 예측 결과와 비교하였을 때 상관계수는 0.679에서 0.868로, NSE는 0.421에서 0.743으로 증가하였다. 본 연구를 통해 머신러닝의 적용으로 기존의 하천 유출 예측 방법을 더 향상시킬 수 있을 것으로 판단된다. Streamflow prediction with physically-based, distributed hydrological modeling requires a long accumulation period of multiple, accurate input datasets. A recently developed hydrologic model, WRF-Hydro modeling system has been used in various regions mainly for short-term simulation. In this study, a machine learning algorithm is applied to improve the forecasting capabilities of the WRF-Hydro modeling system, particularly for improving a long term streamflow prediction. First, the WRF-Hydro modeling system is applied to simulate the inflow of Soyang Dam, which is the upper basin of the Han River with model parameters calibrated by the PEST tool. Then, the machine learning algorithm using LSTM network is trained with time-series of residuals between simulated and observed streamflow. The trained model is used to improve streamflow prediction of the WRF-Hydro modeling system. The 3-hourly inflows of the Soyang Dam are simulated for 2018 and the results are validated in comparison to observation data. From the result, PEST parameter optimization improved 10-day simulation, but showed limited improvement for a 1-year simulation. With applying a machine learning algorithm, the simulation result was improved significantly for both 10-day and 1-year simulations. In comparison to the simulation result where machine learning has not applied, the correlation coefficient value increased from 0.679 to 0.868 and NSE value from 0.421 to 0.743. This study suggests that the machine learning could further improve streamflow prediction that is already constrained by a conventional optimization technique.

라인홀드.니이버의 윤리사상

조경우 嶺南大學校 1983 국내석사