Real-time emission prediction for ship engine using stacked time-series learning: a transformer-XGBoost hybrid...

Area of Science:

• Environmental Science
• Marine Engineering
• Data Science

Background:

• Maritime transportation is crucial for global trade but contributes significantly to air pollution and global warming through vessel diesel engine emissions.
• Current emission inventory methods (e.g., IMO, EEA) lack accuracy in reflecting dynamic engine operational conditions.

Purpose of the Study:

• To develop a hybrid prediction model integrating a time-series forecasting transformer and XGBoost for accurate, real-time vessel emission prediction.
• To improve upon the limitations of existing emission inventory methods in the maritime sector.

Main Methods:

• Utilized a hybrid model combining a time-series forecasting transformer and XGBoost.
• Employed least absolute shrinkage and selection operator (LASSO) regression for key variable selection.
• Trained and validated the model using real-time engine operational data.

Main Results:

• Achieved significant reductions in prediction errors: 34-40% for CO2, 45-47% for CO, and 40-48% for NOx compared to conventional methods.
• Demonstrated stable performance for variable pollutants, with ~85% of CO2 predictions within a ±5% error range.
• Identified limitations related to CO variability and controllable pitch propeller effects.

Conclusions:

• The proposed hybrid model offers a scalable, real-time framework for predicting vessel emissions.
• This framework enables enhanced air pollution monitoring and supports effective emission reduction strategies in the maritime industry.