主管:中国科学院
主办:中国优选法统筹法与经济数学研究会
   中国科学院科技战略咨询研究院

Chinese Journal of Management Science ›› 2026, Vol. 34 ›› Issue (10): 372-385.doi: 10.16381/j.cnki.issn1003-207x.2024.2065

Previous Articles    

Prediction Method of Industrial Carbon Emissions under Multiple Driving Factors and Small Samples

Xiaofan Lai1, Zeyu Zhang1, Xiutian Shi2(), Weiguo Zhang1   

  1. 1.College of Management,Shenzhen University,Shenzhen 518000,China
    2.School of Economics & Management,Nanjing University of Science & Technology,Nanjing 210094,China
  • Received:2024-11-13 Revised:2025-04-03 Online:2026-10-25 Published:2026-10-09
  • Contact: Xiutian Shi E-mail:xtshi@njust.edu.cn

Abstract:

Accurately predicting the total carbon emissions of industries and identifying the key driving factors of carbon emissions are crucial for decision-making in carbon emission control, significantly contributing to China's “dual carbon” goal. However, the accounting process of carbon emission data in China started relatively late and is characterized by small samples, multiple driving factors, high volatility, and non-monotonic patterns, which makes precise predictions particularly challenging. To enhance both the feature selection and generalization capabilities of prediction models under conditions involving small samples and multiple driving factors, a novel method for predicting industrial carbon emissions is proposed.This method mainly consists of three modules as follows:(1) Feature Engineering Module: It uses the Deep Q-Network (DQN) learning method to analyze the correlation between higher-order cross features and carbon emissions sequences, extracting the driving factor sequence that best characterizes the industrial carbon emission trend, thereby avoiding interference from redundant factors during the information integration of the prediction model.(2)Parameter Optimization Module: The Sparrow Search Algorithm (SSA) is employed to optimize the penalty parameter c and kernel parameter g of Support Vector Regression (SVR), improving the model training performance and enhancing its generalization capability.(3)Prediction Module: Based on the state features St output by DQN with St={f1,…, fn} representing the state of the n driving factors fi at the t-th iteration, and based on the optimal parameters c and g provided by SSA, the widely used SVR model is selected to train and predict the industrial carbon emission sequence. The resulting model based on SVR is able to achieve accurate predictions for industrial carbon emissions.The data used in this study are from China's construction and transportation industries from 2005 to 2020, and are sourced from the National Bureau of Statistics of China, Carbon Emission Accounts and Datasets (CEADS) and the China Building Energy Conservation Association. During the numerical experiment, cross-validation and multi-step time series forecasting are used to conduct 2-step, 3-step, and 5-step prediction experiments, resulting in a total of 18 experiments. The numerical results reveal that the proposed predicting method surpasses traditional predicting methods, such as LSTM, ELM, GRU, BP, etc., demonstrating greater robustness and predictive ability, and achieving more precise and effective forecasts of industrial carbon emission trends characterized by small samples and multiple driving factors.

Key words: small samples, multiple driving factors, industrial carbon emissions forecast, feature selection, parameter optimization

CLC Number: