Unraveling Green Innovation Drivers in High Carbon-Lock-in Process Industries: An Interpretable Machine Learning Approach with Temporal Validation

Abstract The petrochemical industry, characterized by severe carbon lock-in and high asset specificity, is one of the most challenging sectors for green transition. Existing studies predominantly rely on linear econometric models that struggle to capture nonlinear synergies and lack industry-specific evidence. Using 380 firm-year observations from Chinese A-share listed petrochemical enterprises (2014–2024), we construct a five-dimensional driver system with 20 indicators grounded in institutional theory and the resource-based view, employing an interpretable XGBoost–SHAP framework. This study makes three key contributions: (1) revealing the conditional boundary of institutional pressure under high asset specificity, showing that only command-and-control regulation forms positive synergies with internal resources; (2) adopting expanding-window rolling validation instead of random partitioning to test temporal generalization, and introducing Kendall’s τ for the stability of SHAP feature importance rankings; and (3) uncovering single-threshold patterns across all core drivers in a high carbon-lock-in industry, providing micro-level evidence for optimizing regulatory instrument mixes. The model achieves a three-year average out-of-sample R² of 0.8352. Internal resource endowments—particularly firm size and R&D investment—dominate green innovation performance, while command-and-control regulation outperforms market-based instruments. These findings offer actionable insights for advancing the low-carbon transition of high-emission process industries, aligned with SDG 9 and SDG 13.