Analyzing different energy transition patterns and their drivers using clustering and machine learning models: unveiling the role of digital technological innovation

This study develops a dual-indicator framework combining renewable energy share (RE) and carbon emissions intensity (CE) to characterize heterogeneous energy transition patterns, addressing limitations of single-metric approaches that mask trade-offs between energy transformation and carbon emission. Employing a two-stage machine learning (ML) pipeline on city-year observations from 273 Chinese cities (2010−2021), we identify three distinct patterns through K-means clustering—Low (LET), Medium (MET), and High (HET) energy transition—each validated by significant inter-cluster distributional differences. Subsequently, an XGBoost classifier coupled with SHAP decomposition quantifies model-based feature contributions exhibiting pronounced nonlinear characteristics and context-dependencies. Network penetration emerges as the dominant yet contingent contributor, shifting from strongly inhibitory in the LET pattern to facilitative in the HET pattern. Digital technology innovation (DTI) exhibits threshold behavior, negatively associated with transition probability below the threshold but positively associated above. Against prior expectations, both fiscal expenditure and scientific investment exhibit negative SHAP contributions. These findings question the assumption of uniform technological benefits and highlight the risks of aggregation bias in urban-level analyses. The framework advances energy transition theory by quantifying previously qualitative concepts of system complementarities, offering actionable evidence for precision decarbonization strategies amid China's carbon neutrality commitments.