基于库普曼理论的中国工业碳排放预测建模  

Predictive modeling of carbon emissions in Chinese industry based on Koopman theory

作　　者：董炜 车金星[1] DONG Wei;CHE Jinxing(School of Science,Nanchang Institute of Technology,Nanchang 330099,China)

机构地区：[1]南昌工程学院理学院,江西南昌330099

出　　处：《南昌工程学院学报》2025年第1期107-112,共6页Journal of Nanchang Institute of Technology

基　　金：国家自然科学基金资助项目(72471108);江西省高等学校教学改革研究课题(JXJG-23-18-21)。

摘　　要：为了提高中国工业碳排放预测的精确度,以支持碳排放管理和政策制定,针对碳排放时间序列的非平稳特性,提出了一种基于库普曼理论的新型预测模型。该模型利用傅里叶滤波器将复杂的非平稳序列分解为时变和时不变成分,并结合多层感知机构建编码器和解码器,以捕捉数据的非线性特征。实验结果表明,与主流的LSTM和GRU模型相比,本模型在预测精度上具有显著优势,RMSE降低至0.054,MAPE降低至3.443%,R 2提高到0.957。尽管在参数数量和计算时间上有所增加,但本模型预测性能的提升可为工业碳排放趋势分析和减排政策制定提供更可靠的数据支持。This study aims to improve the accuracy of carbon emission prediction in China’s industrial sector to support carbon management and policy-making.In response to the non-stationary characteristics of carbon emissions time series,this paper proposes a novel prediction model based on Koopman theory.The model utilizes Fourier filters to decompose complex non-stationary sequences into time-varying and time-invariant components and employs a multilayer perceptron to build an encoder-decoder structure to capture the nonlinear features of the data.Experimental results show that compared to traditional LSTM and GRU models,the proposed model exhibits significant advantages in prediction accuracy,reducing RMSE to 0.054,lowering MAPE to 3.443%,and improving R 2 to 0.957.Although the model introduces an increase in parameter count and computation time,the improvement in predictive performance provides more reliable data support for industrial carbon emission trend analysis and emission reduction policy formulation.

关 键 词：库普曼算子 多层感知机 短期碳排放预测 傅里叶分析 

分 类 号：X321[环境科学与工程—环境工程]