上海交通大学学报

上海交通大学学报

• •    下一篇

基于卷积神经网络与轻量级梯度提升树组合模型的电力行业短期以电折碳方法(网络首发)

  

  1. 1. 南方电网能源发展研究院；2. 清华四川能源互联网研究院

A Short-Term Carbon Emission Accounting Method Using Electricity Data Based on Convolutional Neural Networks and Light Gradient Boosting Machine Combined Model in Power Industry

  1. 1. Energy Research Institute of China Southern Power Grid Co., Ltd, Guangzhou 510700, China; 2. Tsinghua Sichuan Energy Internet Research Institute, Chengdu 610213, China

PDF (PC)

151

摘要： 电力行业是碳排放的重点控排行业，准确、实时的计量电力行业碳排放是支撑其降碳减排的基础。目前，电力行业的碳排放计量主要基于实测法或核算法，难以很好的兼顾低计量成本与实时计量能力。为此，本文充分考虑电力行业良好的电力数据基础，充分挖掘电-碳间的相关关系，以电力历史数据为基础，基于机器学习方法提出了一种电力行业短期以电折碳方法，以实时估算电力行业短期碳排放情况。该方法使用卷积神经网络（Convolutional Neural Networks， CNN）进行特征提取，并采用轻量级梯度提升树算法（Light Gradient Boosting Machine， LightGBM）开展基于特征提取值的碳排放测算。此外，为了提升模型的泛化能力和鲁棒性，本文在模型训练中采用了K折交叉验证技术，在模型参数优化过程中采用了网格搜索技术。最后，为了验证所提模型的有效性，本文对比了所提模型和其他机器学习模型在同等数据集划分条件下分别基于日度数据集与小时数据集中进行训练的效果，结果表明所提模型在效果评估和测算值与目标值分布分析中均优于其它模型，能够较好的反映电力行业的短期碳排放情况。

关键词: 以电折碳, 卷积神经网络, 轻量级梯度提升树算法, 碳排放, 机器学习, 组合模型

Abstract: The electric power industry is a key sector for carbon emission control. Accurate and real-time accounting of carbon emissions in the power industry is the basis for supporting the carbon reduction of power industry. At present, the measurement of carbon emissions in the power industry is mainly based on the actual measurement method or the accounting method, which is difficult to balance low measurement costs and real-time measurement capabilities. To this end, this paper fully considers the good power data foundation of the power industry, fully explores the correlation between electricity and carbon, and proposes a short-term electricity-to-carbon method in the power industry based on machine learning methods based on historical data of electricity. In this method, feature extraction is conducted with Convolutional Neural Networks (CNNs), and the Light Gradient Boosting Machine (LightGBM) is used for carbon emission estimation based on extracted features. Moreover, K-Fold cross-validation is used in model training, and parameter optimization is performed using grid search to enhance the model's generalization capability and robustness. To validate the effectiveness of the proposed model, other machine learning models were tested under the same data segmentation on daily and hourly data sets, for model performance evaluation. The results indicate that the proposed model outperforms other models in both performance evaluation and the consistency between estimated and target values.

Key words: electric folding carbon, convolutional neural networks, light gradient boosting machine, carbon emissions, machine learning, combined model

中图分类号: