머신러닝을 활용한 항만 예인선 작업의 연료소모와 온실가스 배출 요인 분석

다국어 초록 (Multilingual Abstract)

Harbor tugboats play a critical role in assisting large vessels during berthing, unberthing, and port maneuvering operations, which involve repeated low-speed operation with high power output. These operational characteristics result in relatively high fuel consumption and greenhouse gas (GHG) emissions compared to vessel size. Despite this, harbor tugboats have been excluded from the International Maritime Organization (IMO) GHG reduction regulations, and quantitative studies based on real operational data remain limited. This study therefore analyzes fuel consumption and GHG emission factors of harbor tugboats using actual operational data. Fuel consumption (L) was set as the dependent variable, with workload (DWT), sailing distance (km), and wind speed (m/s) as independent variables. To address small-sample and nonlinear characteristics, bootstrapping and SMOTER-based resampling were applied to regression analysis, while XGBoost, tuned XGBoost, and Random Forest models were employed for machine learning analysis. Model performance was evaluated using RMSE, MAE, and MAPE, and variable influence was interpreted using standardized regression coefficients (β) and SHAP analysis. Model stability was assessed using
Leave-One-Out Cross-Validation (LOOCV) The results indicate substantial differences in prediction errors between regression and machine learning models, preventing direct performance comparison and necessitating independent interpretation. Machine learning models generally exhibited lower prediction errors,
while the XGBoost model showed notable sensitivity to hyperparameter tuning. LOOCV results emphasized the relative influence of navigational characteristics rather than predictive accuracy. Sailing distance consistently emerged as the dominant factor, indicating that tugboat fuel consumption is primarily governed by cumulative operational characteristics rather than instantaneous loads, whereas the influence of workload (DWT) varied depending on the resampling method.
This study confirms the applicability of machine learning methods for analyzing harbor tugboat fuel consumption and GHG emission factors using real operational data, and highlights the importance of methodological considerations under small-sample and nonlinear conditions. Future research separating generator and main engine operations may further clarify the respective influences of workload and sailing distance.