A STUDY OF PROVINCIAL CARBON EMISSIONS IN CHINA'S CONSTRUCTION INDUSTRY USING THE STIRPAT AND DEEP LEARNING MODEL

Weiming Xu, Chrystal Zhang, Fuji Ren, Weiqiang Liu, Weiwei WangApplied Ecology and Environmental Research 2025 · DOI

Abstract

The construction industry is a significant contributor to energy consumption and carbon emissions, making its decarbonization essential for achieving the "dual carbon" goals. However, existing studies often focus on national or single-region analyses and lack a systematic understanding of provincial-level disparities in emissions and peaking pathways. To address this challenge, this study first employs an extended STIRPAT (Stochastic Impacts by Regression on Population, Affluence, and Technology) model to identify nine key variables influencing carbon emissions, using data from the construction industry across 30 Chinese provinces from 1997 to 2021. Three scenarios—baseline, green development, and high-speed development—are developed for analysis. A novel forecasting model integrating Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and attention mechanisms is then proposed to predict the carbon emission trends and peaking timelines of China's construction sector from 2022 to 2035. The results indicate that the proposed model achieves higher prediction accuracy than traditional methods. Across all scenarios, carbon peaking targets are projected to be reached before 2030 in all provinces. The green development scenario results in faster emission reductions and earlier peaking, while the high-speed development scenario delays peaking in some central and western provinces, highlighting greater mitigation potential and spatial heterogeneity. Finally, policy recommendations based on the findings provide scientific guidance for the low-carbon transition of the construction industry and the formulation of province-specific emission reduction strategies by government authorities.

Introduction

China has committed to ambitious "dual carbon" goals: peaking carbon emissions by 2030 and achieving carbon neutrality by 2060. As a cornerstone of national development, the construction industry accounts for over 50% of the country's total lifecycle carbon emissions. Consequently, decarbonizing this sector is critical for national climate targets.

Existing research often treats China as a monolith or focuses on isolated regions, failing to account for the vast economic and geographic disparities between provinces. Furthermore, traditional forecasting models frequently struggle with the complex, non-linear, and "noisy" nature of construction data, which is influenced by diverse factors ranging from urbanization rates to fluctuating material costs. This study addresses these gaps by providing a granular, province-level analysis using a sophisticated hybrid deep learning approach.

Methodology

Feature Selection: Extended STIRPAT Model

To identify the primary drivers of carbon emissions, the researchers utilized an extended STIRPAT model. This framework allowed for the selection of nine critical variables, including Gross Domestic Product (GDP), urbanization rate, energy intensity, and industry-specific metrics such as completed building area. A correlation analysis was performed to validate the relationships between these variables and carbon intensity.

Correlation analysis of key variables influencing carbon emissions

The Hybrid CNN-GRU-LSTM-Attention Model

The core of the forecasting framework is a multi-layered deep learning architecture designed to handle both spatial and temporal complexities:

  1. CNN (Convolutional Neural Networks): These layers are employed to extract local spatial features and patterns from the input data, effectively identifying structural relationships within the provincial datasets.
  2. GRU (Gated Recurrent Units) & LSTM (Long Short-Term Memory): These components capture temporal dependencies. While GRU handles short-term fluctuations, LSTM manages long-term sequence dependencies, mitigating the "vanishing gradient" problem common in standard recurrent networks.
  3. Attention Mechanism: This serves as the "secret sauce," assigning higher weights to the most critical historical data points, which significantly enhances the model's predictive precision.
Architecture of the Convolutional Neural Network (CNN) layer used for spatial feature extraction

Data and Scenarios

The study utilized a comprehensive 25-year dataset (1997–2021) covering 30 Chinese provinces. Data was meticulously sourced from official national yearbooks regarding population, energy consumption, and construction industry output. The dataset was divided into a 20-year training set (600 samples) and a 5-year testing set (150 samples).

To explore potential futures, three scenarios were simulated for the 2022–2035 period:

  • Baseline Scenario: Continuation of current trends and policies.
  • Green Development Scenario: Aggressive adoption of green technologies and stringent emission standards.
  • High-Speed Development Scenario: Prioritization of rapid economic growth and infrastructure expansion.

Results and Discussion

Model Performance

The hybrid CNN-GRU-LSTM-Attention model demonstrated exceptional accuracy, achieving an R² value of 0.993. This significantly outperformed standalone models; for instance, a basic LSTM model yielded a Mean Absolute Percentage Error (MAPE) of 50.65%, whereas the hybrid model reduced this error to just 4.73%.

Peaking Timelines and Spatial Heterogeneity

The analysis revealed significant spatial heterogeneity across China. Economically developed regions in the East and North, such as Beijing and Shanghai, have largely already peaked and entered a decline phase. In contrast, Central and Western provinces like Sichuan and Guangxi remain in an "expansion phase" due to ongoing infrastructure needs.

Under the Baseline and Green scenarios, national construction emissions were estimated to have peaked in 2019 at approximately 140 million tons. However, the High-Speed scenario suggests a potential delay, with a higher peak of 143 million tons occurring in 2026. The Green Development path is particularly effective, as it could enable all provinces to reach their targets as early as 2025.

Conclusion

This study highlights that while a national carbon peak before 2030 is achievable for the construction industry, "green development" pathways are essential for accelerating the transition, especially in rapidly urbanizing regions. The findings suggest that Central and Western provinces offer the highest mitigation potential, where implementing green building standards now will yield the greatest long-term returns.

Future work should address the limitations of data uncertainty and the potential for "black swan" policy shifts. Additionally, incorporating spatial econometric analyses to model the interactions and potential "carbon leakage" between neighboring provinces would further refine the accuracy of provincial roadmaps.

Summary updated

Abstract

The construction industry is a significant contributor to energy consumption and carbon emissions, making its decarbonization essential for achieving the "dual carbon" goals.However, existing studies often focus on national or single-region analyses and lack a systematic understanding of provinciallevel disparities in emissions and peaking pathways.To address this challenge, this study first employs an extended STIRPAT (Stochastic Impacts by Regression on Population, Affluence, and Technology) model to identify nine key variables influencing carbon emissions, using data from the construction industry across 30 Chinese provinces from 1997 to 2021.Three scenarios-baseline, green development, and high-speed development-are developed for analysis.A novel forecasting model integrating Convolutional Neural Networks (CNN), Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM) networks, and attention mechanisms is then proposed to predict the carbon emission trends and peaking timelines of China's construction sector from 2022 to 2035.The results indicate that the proposed model achieves higher prediction accuracy than traditional methods.Across all scenarios, carbon peaking targets are projected to be reached before 2030 in all provinces.The green development scenario results in faster emission reductions and earlier peaking, while the high-speed development scenario delays peaking in some central and western provinces, highlighting greater mitigation potential and spatial heterogeneity.Finally, policy recommendations based on the findings provide scientific guidance for the low-carbon transition of the construction industry and the formulation of province-specific emission reduction strategies by government authorities.
BibTeX
@inproceedings{xu2025study,
  author = {Weiming Xu and Chrystal Zhang and Fuji Ren and Weiqiang Liu and Weiwei Wang},
  title = {{A STUDY OF PROVINCIAL CARBON EMISSIONS IN CHINA'S CONSTRUCTION INDUSTRY USING THE STIRPAT AND DEEP LEARNING MODEL}},
  booktitle = {Applied Ecology and Environmental Research},
  year = {2025},
  doi = {10.15666/aeer/2306\_1075110774},
}
Figure