基于符号回归的中国碳排放影响因素识别

Identification of factors influencing China's carbon emissions based on symbolic regression

关键词:  碳排放  符号回归  遗传算法  区域差异

Key words:  carbon emission  symbolic regression  genetic algorithm  regional heterogeneity

摘要:

在“双碳”倡议背景下，科学识别碳排放的驱动因素已成为环境经济学的重要议题。文章以中国30个省级行政区（不含西藏及港澳台地区）为研究对象，基于2008—2022年的省级面板数据，选取人口规模、经济发展水平、技术投入、人力资本、外商直接投资、产业结构及能源结构等变量，引入符号回归模型进行分析。该方法以遗传编程为核心搜索机制，通过迭代演化自动生成能够刻画变量非线性关联特征的数学表达式，从而识别不同省级行政区碳排放相关变量的主要结构特征。研究结果表明，中国省域碳排放具有显著的区域异质性与非线性特征：（1）在主要驱动因素分布方面，国内生产总值（GDP）是10个省级行政区的核心驱动变量，人口规模在6个省级行政区中作用显著，外国直接投资在5个省级行政区中具有重要影响，能源消费结构在3个省级行政区中发挥关键作用，技术创新水平、人力资本水平和产业结构则分别在两个省级行政区中表现突出。体现了不同地区碳排放驱动因素的多样性。（2）符号回归模型有效识别出人口主导型、经济主导型、技术主导型、外资主导型、产业主导型与能源主导型等多类地区。其中，以GDP为核心因素的省级行政区人均碳排放水平最低，而以外商投资水平为核心因素的省级行政区碳排放水平最高；能源结构主导型地区整体经济水平相对偏低，但减排潜力较大。基于上述结论，可为不同地区因地制宜制定低碳发展政策提供参考，并为进一步开展多尺度碳排放驱动机制研究提供思路。

Abstract:

Background, aim, and scope Against the backdrop of the “Dual-Carbon” initiative, China has committed to peaking its carbon dioxide emissions before 2030 and achieving carbon neutrality by 2060. However, due to China’s vast geographical expanse, disparate resource endowments, and varying stages of regional development, the driving forces behind carbon emissions are highly complex. They exhibit strong regional heterogeneity and nonlinear dynamics that traditional linear models often fail to capture. The primary aim of this study is to systematically decode these complex, nonlinear driving mechanisms at the regional level, isolating the dominant structural drivers of provincial carbon emissions to inform targeted mitigation strategies. Materials and methods The scope of this empirical research encompasses a comprehensive analysis of 30 Chinese provincial-level administrative regions (excluding Xizang, Hong Kong, Macao and Taiwan regions), utilizing provincial panel data spanning 2008 to 2022. To fully reflect the multidimensional characteristics of regional carbon emissions, the evaluation framework constructed in this study incorporates seven explanatory variables: population size, economic development level, technological input, human capital, foreign direct investment, industrial structure and energy consumption structure. This study adopts the symbolic regression method for analysis. By iteratively generating mathematical expressions, it automatically excavates the inherent real action laws of data and captures latent nonlinear correlations and complex interaction effects among variables. Results The empirical results show that provincial carbon emissions in China exhibit profound regional heterogeneity and pronounced nonlinear features. First, regarding the spatial distribution of primary factors, the model successfully isolates each provincial-level administrative region’s dominant driver: GDP is identified as the core factor in 10 provincial-level administrative regions; population size is significant in 6 provincial-level administrative regions; FDI plays a crucial role in 5 provincial-level administrative regions; and energy consumption structure is critical in 3 provincial-level administrative regions. Furthermore, technological innovation, human capital, and industrial structure each demonstrate dominant effects in two provincial-level administrative regions. Second, based on these isolated drivers, the study categorizes the regions into four distinct typologies: economic-dominant, population-dominant, FDI-dominant, and energy-dominant types. Comparatively, GDP-dominant provincial-level administrative regions exhibit the lowest overall carbon emissions, FDI-dominant provincial-level administrative regions exhibit the highest emission levels, and energy-dominant provincial-level administrative regions display the most substantial potential for carbon mitigation. Discussion The findings of this study reveal the structural driving factors of carbon emissions across Chinese provincial-level administrative regions from multiple dimensions. Provincial-level administrative regions dominated by economic growth record the lowest total carbon emissions, and such regions have likely passed the inflection point of the Environmental Kuznets Curve. Economic expansion facilitates industrial structure optimization and generates technological spillover effects, promoting the shift toward high-value-added, low-carbon tertiary industries. On the contrary, provincial-level administrative regions dominated by foreign direct investment sustain persistently high carbon emissions, which serves as a regional manifestation of the pollution haven hypothesis. Massive foreign capital flows into energy-intensive and low-end manufacturing sectors, shifting carbon emission burdens to local regions. Energy-dominant provincial-level administrative regions boast enormous emission reduction potential. Having long relied on fossil fuels for development, these regions are trapped in fossil fuel lock-in effects, yet they constitute core areas for developing renewable energy and delivering the largest national carbon emission cuts. From a methodological perspective, the symbolic regression model presents remarkable analytical strengths. It automatically conducts rigorous variable selection for high-dimensional data and generates structurally sound, differentiated mathematical expressions tailored to each provincial-level administrative region to accommodate complex regional disparities. Moreover, it captures interactive relationships that conventional linear models fail to identify. Conclusions In conclusion, symbolic regression can help identify the carbon emission driving typologies of each provincial-level administrative region and clarify the unique interaction patterns of core factors in corresponding regions. The driving mechanisms of carbon emissions in China are marked by distinct regional features, diversified structural characteristics and prominent nonlinearity. Gross domestic product, population size, foreign direct investment and energy consumption structure are the nationwide core influencing factors, and the dominant effect of each factor shows significant geographical differentiation. Recommendations and perspectives This study thus puts forward differentiated emission reduction policy recommendations. Provincial-level administrative regions with large populations should focus on improving energy utilization efficiency and optimizing the layout of low-carbon infrastructure; economically developed provincial-level administrative regions need to continuously advance green and low-carbon transformation to achieve high-quality development. It is essential to guide foreign direct investment toward low-carbon industries. Meanwhile, all regions should simultaneously strengthen technological innovation and human capital cultivation to accelerate the implementation and popularization of low-carbon technologies. Future research can further expand the system of carbon emission influencing factors, introduce various machine learning algorithms and add model constraints, and deeply analyze the driving mechanisms of regional carbon emissions from multiple dimensions.