대청호의 계절별 CO2 NAF 산정 및 기계 학습 모델의 적용성 평가

다국어 초록 (Multilingual Abstract)

Quantitative estimates of the CO2 Net Atmospheric Flux (NAF) of lakes and reservoirs in various climatic regions have been presented based on the thermodynamic equilibrium theory of carbonate system, but there is still great uncertainty in global estimates. This is because the CO2 NAFs are largely dependent on the partial pressure of carbon dioxide (PCO2) in the water, but the PCO2 measurement data are mostly rare. The purpose of this study was to estimate the CO2 NAF of Daecheong Reservoir (Korea) in 2012 and 2013 considering the uncertainty of the PCO2 estimation method using the filed data collected at various surface waters in Geum River and Saemangeum basin of Korea. In addition, multiple linear regression models (MLR) and machine learning models (RF and ANN) were used to identify the major environmental factors that determine daily PCO2 variations in Daecheong Reservoir, and developed the NAF prediction models with selected input variables. This result showed that pH, Alk, and DIC measurement data are thermodynamically satisfactory within the carbonate system, although calculated PCO2 is highly sensitive to the accuracy of pH measurements, particularly at low pH. Daecheong Reservoir was found to be the source of atmospheric CO2 emission, and the NAFs in 2012 and 2013 were 2,590 and 771 mg CO2 m-2d-1, respectively. A stepwise multiple regression model selected five independent variables (WT, pH, Alk, Chl-a, Uw) for the parsimonious model. The R2 values of MLR, RF, and ANN for the estimated NAF were 0.699, 0.975, 0.997. The RF and ANN models showed much enhanced performance in the estimation of the high NAF values, while MLR model significantly underestimated them. A cross validation with 10-fold random samplings was applied to evaluate the performance of three models, and indicated that the ANN model is best, and followed by RF and MLR models.

목차 (Table of Contents)

• Ⅰ. 서 론 1
• Ⅱ. 이론고찰 5
• 2.1 탄산염 시스템 5
• 2.1.1 닫힌계 (Closed System) 5
• 2.1.2 열린계 (Open System) 7
• 2.2 기체 전달 속도 9
• 2.3 다중회귀 및 기계 학습 모델 10
• 2.3.1 다중회귀(MLR)모델 10
• 2.3.2 랜덤포레스트(Random forest, RF) 12
• 2.3.3 인공신경망(Artificial Neural Network, ANN) 14
• Ⅲ. 연구방법 및 범위 17
• 3.1 연구 대상지역 17
• 3.2 현장조사 및 분석방법 20
• 3.3 CO2 배출량 산정 22
• 3.4 기계학습 모델 적용 절차 24
• 3.4.1 RF 모델 25
• 3.4.2 ANN 모델 27
• 3.5 모델 검정 및 평가 29
• Ⅳ. 연구 결과 31
• 4.1 PCO2 산정 불확도 평가 31
• 4.1.1 실험결과 기술통계 분석 31
• 4.1.2 교차상관분석 34
• 4.1.3 PCO2 산정 불확도 비교 36
• 4.1.4 민감도 분석 38
• 4.2 CO2 배출량 산정 및 기계 학습 모델 개발 39
• 4.2.1 입력변수 및 영향변수의 특성 39
• 4.2.2 DIC와 k값 산정결과 44
• 4.2.3 CO2 NAF산정 결과 47
• 4.2.4 NAF와 영향인자의 회귀분석 49
• 4.2.5 MLR 모델 모의 결과 52
• 4.2.6 랜덤포레스트(RF) 모델 모의 결과 53
• 4.2.7 인공신경망(ANN) 모델 모의 결과 55
• 4.2.8 교차 검정 결과 57
• 4.2.9 계절별 CO2 NAF 산정 성능 평가 59
• 4.2.10 선행연구결과와 비교고찰 62
• Ⅴ. 결 론 63
• 참고 문헌 66