Comparative study of machine learning methods for carbon metering in power generation enterprises

Area of Science:

• Environmental Science
• Energy Engineering
• Data Science

Background:

• Global carbon neutrality goals necessitate accurate emission monitoring for power generation.
• Traditional carbon accounting methods are limited in accuracy, adaptability, and cost.
• Machine learning (ML) presents a viable alternative for precise carbon emission prediction.

Purpose of the Study:

• To develop and compare ML models for predicting carbon emissions from a coal-fired power plant.
• To identify key predictors influencing carbon emissions.
• To provide an efficient and accurate method for carbon emission calculation.

Main Methods:

• Utilized operational data from a coal-fired power plant (Jan 1 - Nov 30, 2024).
• Applied a hybrid feature selection strategy to identify 18 key parameters.
• Developed and optimized multiple linear regression, XGBoost, and LSTM models via hyperparameter tuning.

Main Results:

• Identified power generation, energy structure, operating time, and load rate as key predictors.
• Achieved a prediction accuracy of 90.39% using the XGBoost model.
• Demonstrated XGBoost's capability in capturing complex emission patterns.

Conclusions:

• The XGBoost model offers a highly accurate and efficient method for carbon emission prediction in power generation.
• This research provides a scientific basis for power enterprises' carbon management and emission reduction decisions.
• ML-driven carbon accounting supports the transition towards carbon neutrality.