【作者】 王颖；

【导师】 程桂芳；

【作者基本信息】 郑州大学 ， 应用统计（专业学位）， 2025， 硕士

【摘要】 山西省作为我国重要的能源生产基地,长期以来面临着空气质量与碳排放治理的双重压力。碳排放预测作为“双碳”工作的重心,近年来取得了显著进展。山西省碳排放量具备“煤电依赖性强”和“气象敏感性高”的双重特征,使用传统统计模型进行预测存在精度不高、优化效率低的问题,准确预测其排放量对区域低碳转型至关重要。论文从多维度引入碳排放量影响因素,并将传统统计分析模型和神经网络深度学习方法相结合,对山西省碳排放量进行预测分析。论文首先使用随机森林、极限梯度提升算法和轻量级梯度提升机三种特征选择方法综合选择出碳排放量影响因素,并在此基础上使用SHAP值可视化和双重机器学习因果推断框架对变量的影响机制进行解释;接下来建立季节性自回归积分移动平均(SARIMA)模型、长短期记忆神经网络(LSTM)和门控循环单元(GRU)等单一模型预测碳排放量。结果表明GRU模型的预测效果最好,决定系数R~2为0.8868,均方误差MSE为0.0018,这说明GRU模型在处理碳排放量这类中等复杂度的时间序列时表现较好。为了进一步提升模型的预测性能,论文在单一模型基础上将SARIMA模型与GRU模型相结合,采用串联方式构建SARIMA-GRU组合预测模型,决定系数R~2提升至0.9211,均方误差MSE降为0.0017。由于组合模型在训练过程中损失函数存在震荡现象,论文引入蚁群优化算法(ACO)和粒子群优化算法(PSO)构建对比实验对模型参数调优,结果表明SARIMA-PSO-GRU模型的预测效果更优,R~2进一步提升至0.9344,MSE降至0.0012。论文在SARIMA-PSO-GRU模型基础上引入多头注意力机制(MHA)增强对关键信息的捕获能力,同时添加SARIMA-Transformer预测模型丰富模型对比视角,结果显示SARIMA-PSO-MHA-GRU模型的预测效果最优,其R~2达到0.9514,MSE降至0.0011。这充分表明组合模型结合了传统时间序列模型、机器学习技术和深度学习模型的特点,能够较好地拟合碳排放量复杂变化规律。更多还原

【Abstract】 As an important energy production base in China,Shanxi Province has long been faced with the dual pressures of air quality and carbon emission governance.Carbon emission prediction,as the focus of the“dual carbon”work,has made remarkable progress in recent years.The carbon emissions in Shanxi Province have the dual characteristics of“strong dependence on coal-fired power”and“high sensitivity to meteorological conditions”.When using traditional statistical models for prediction,there are problems such as low accuracy and low optimization efficiency.Accurately predicting its emissions is crucial for the regional low-carbon transformation.This thesis introduces the influencing factors of carbon emissions from multiple dimensions,combines the traditional statistical analysis model with the deep learning method of neural networks,and conducts a predictive analysis of the carbon emissions in Shanxi Province.Firstly,this thesis uses three feature selection methods,namely Random Forest,Extreme Gradient Boosting algorithm,and Light Gradient Boosting Machine,to comprehensively select the influencing factors of carbon emissions.On this basis,the SHAP value visualization and the causal inference framework of double machine learning are used to explain the influence mechanism of variables.Next,single models such as the Seasonal Autoregressive Integrated Moving Average(SARIMA)model,Long Short-Term Memory neural network(LSTM)model,and Gated Recurrent Unit(GRU)model are established to predict carbon emissions.The results show that the GRU model has the best prediction effect,with the coefficient of determination R~2reaching 0.8868 and the Mean Squared Error(MSE)being 0.0018.This indicates that the GRU model performs well in handling time series of medium complexity such as carbon emissions.In order to further improve the prediction performance of the model,this thesis combines the SARIMA model with the GRU model on the basis of the single model,and constructs a SARIMA-GRU combined prediction model in a cascaded manner.The coefficient of determination R~2 is increased to 0.9211,and the Mean Squared Error(MSE)is reduced to 0.0017.Due to the oscillation phenomenon of the loss function during the training process of the combined model,this thesis introduces the Ant Colony Optimization algorithm(ACO)and the Particle Swarm Optimization algorithm(PSO)to construct a comparative experiment for tuning the model parameters.The results show that the prediction effect of the SARIMA-PSO-GRU model is better.The R~2 is further increased to 0.9344,and the MSE is reduced to 0.0012.Based on the SARIMA-PSO-GRU model,this thesis introduces the Multi-Head Attention mechanism(MHA)to enhance the model’s ability to capture key information.At the same time,the SARIMA-Transformer prediction model is added to enrich the perspective of model comparison.The results show that the prediction effect of the SARIMA-PSO-MHA-GRU model is the best,with its R~2 reaching 0.9514 and the MSE decreasing to 0.0011.This indicates that the combined model combines the characteristics of traditional time series models,machine learning techniques,and deep learning models,and can well fit the complex variation laws of carbon emissions.更多还原