华南理工大学学报(自然科学版)

• 环境科学与技术 • 上一篇    下一篇

基于多源数据及机器学习电厂CO2计量与预测研究

孙尊强 田一淳2  苏楠 郑成航3  张振 杨宏旻 高翔3   

  1. 1.国电环境保护研究院有限公司,江苏 南京 210031;

    2.北京低碳清洁能源研究院,北京 102200

    3.浙江大学,浙江 杭州 310027

    4.南京师范大学 能源与机械工程学院,江苏 南京 210042

  • 发布日期:2025-07-01

CO2 Emission Measurement and Forecasting for Power Plants Based on Multi-Source Data and Machine Learning

SUN Zunqiang1  TIAN Yichun2  SU Nan2  ZHENG Chenghang3  ZHANG Zhen4  YANG Hongmin4  GAO Xiang3   

  1. 1. State Power Environmental Protection Research Instituter, Nanjing 210031, Jiangsu, China;

    2. National Institute of Clean and Low-carbon Energy, Beijing 102200, China;

    3. Zhejiang University, Hangzhou 310027, Zhejiang, China;

    4. School of Energy and Mechanical Engineering, Nanjing Normal University, Nanjing 210042, Jiangsu, China

  • Published:2025-07-01

摘要:

准确的CO2排放计量与动态预测对于推动中国的“双碳”目标(碳达峰与碳中和)的实现至关重要。本研究结合了排放因子法、CO2-CEMS连续测量法和机器学习技术,提出了一种基于多源数据融合的CO2排放计量与预测方法,旨在为燃煤电厂提供一个高效且精确的碳排放监测工具。研究通过对比不同方法的碳排放量计算结果,评估了不同机器学习算法(如Adaboost等)的性能,并基于多源数据开发了一个动态预测模型。在国能河北某电厂1#和3#机组上的实验结果表明,排放因子法和CO2-CEMS法的CO2排放量计算相对偏差分别为1.63%和-1.27%,排放因子法和CO2-CEMS连续监测法具有较好的互证性。使用多种机器学习模型(如XGBoost、LightGBM等)与Adaboost模型进行了对比,除了R²和MAPE等常规评价指标,提出一种新的筛选标准,即基于已训练模型再应用至其他机组推到得出的均值偏差xc,用于评估机器学习算法的泛化能力及进一步筛选。结果发现Adaboost在预测精度和稳定性方面表现最佳,并具有更高的泛化能力与鲁棒性。通过优选的Adaboost算法进行的CO2排放动态预测,在训练集和测试集上的R²值均大于0.99,MAPE值低于2%,表明该模型具有很高的预测准确性和稳定性,并且具有极强的泛化能力和鲁棒性。提出的多源数据融合方法不仅有效克服了传统方法在动态变化中的局限性,还能够根据实时数据进行小时级的CO2排放量精准预测。

关键词: CO2, 碳计量, 碳预测, 排放因子法, 机器学习

Abstract:

Accurate CO2 emission measurement and dynamic forecasting are crucial for achieving China's "dual carbon" goals (carbon peak and carbon neutrality). This study integrates the emission factor method, CO2-CEMS, and machine learning technologies to propose a CO2 emission measurement and forecasting method based on multi-source data fusion, aimed at providing an efficient and accurate carbon emission monitoring tool for coal-fired power plants. By comparing the CO2 emission calculation results of different methods, the study evaluates the performance of machine learning algorithms, such as Adaboost, and develops a dynamic forecasting model based on multi-source data. Experimental results show that in Units 1 and 3 of a power plant in Hebei, China, the relative deviations of CO2 emission calculations between the emission factor method and CO2-CEMS were 1.63% and -1.27%, respectively, demonstrating good cross-validation between the two methods. Multiple machine learning models (such as XGBoost, LightGBM, etc.) were compared against the AdaBoost model. In addition to conventional evaluation metrics such as R² and MAPE, a novel selection criterion, namely the mean deviation (xc) obtained by applying trained models to other units, was proposed to assess the generalization capability of machine learning algorithms for further model screening. The results revealed that AdaBoost exhibited superior performance in prediction accuracy and stability, along with higher generalization capability and robustness. The dynamic CO2 emission forecasting using the optimized Adaboost algorithm achieved R² values greater than 0.99 on both the training and testing sets, with a mean absolute percentage error (MAPE) below 2%, indicating high prediction accuracy, stability, generalization ability, and robustness. The proposed multi-source data fusion method not only effectively overcomes the limitations of traditional methods in dynamic scenarios but also enables precise hourly CO2 emission forecasting based on real-time data.

Key words: CO2, carbon measurement, carbon forecasting, emission factor method, machine learning