A review of the application of machine learning in carbon emission assessment studies: prediction optimization and driving factor selection

With growing global concerns over the greenhouse effect and its impacts, developing effective carbon emission prediction models has become an imperative for addressing climate change. Machine learning (ML) excels at extracting potential patterns and correlations from a wide range of datasets, playing an indispensable role in this research field. This study conducted a systematic literature review of 126 peer-reviewed papers applying machine learning to carbon accounting, aiming to synthesize research progress, identify knowledge gaps and propose future directions. Key findings indicate that: single-model prediction frameworks dominate current methodologies, with decision tree-based ML models exhibiting the highest prevalence; energy types and consumption structures are recognized as the most critical determinants of carbon emissions; while existing research predominantly emphasizes economic drivers (e.g., GDP and industrial structure), demographic factors such as population dynamics and aging are anticipated to become focal points due to global demographic shifts. Despite ML's demonstrated advantages in predictive accuracy, pattern identification and computational efficiency, persistent challenges remain. Future endeavors should prioritize: standardization of data formats and interoperability protocols, pen-access sharing of datasets and analytical code and fostering interdisciplinary collaboration. This review establishes a systematic framework for advancing carbon accounting research, offering insights to catalyze cross-disciplinary innovation in the field.