How to solve the small sample problem in the building sector carbon emission peak simulation: a machine learning based multi-stage prediction approach

Accurately predicting Building Sector Carbon Emissions (BSCEs) is crucial for implementing effective carbon emissions reduction policies. Nevertheless, significant regional heterogeneity and limited BSCEs data pose major challenges for peak simulation. Existing studies rarely quantify the priority of driving factors across regions or address the small sample issue in machine learning-based BSCEs prediction. To address the challenges mentioned above and fill the research gaps, a multi-stage hybrid BSCEs prediction approach based on machine learning method was developed. The Random Forest method was employed to identify driving factors. The Time-series Generative Adversarial Networks (TimeGAN) method was utilized to augment limited sample data, while the Long Short-Term Memory (LSTM) combined with Monte Carlo Simulation was used to predict BSCEs dynamically. A case study of 30 Chinese provinces evaluated the effectiveness of the proposed approach. The research results show that: a) The driving factors of BSCEs vary significantly among different provinces in China. Energy intensity and energy structure are widely driving factors for most provinces. Government green investment is only a representative factor for Liaoning and Jilin. b) The TimeGAN-LSTM model outperforms the original LSTM model, as evidenced by an average 25 % increase in prediction accuracy on the validation set. c) Empirical results show that eight provinces can achieve the peak of BSCEs before 2030 under three scenarios, while three provinces fail to achieve the goal of BSCEs peak under three scenarios. This study has made an innovative advancement in the methodological framework for BSCE prediction and facilitated the development of region-specific emission reduction policies.