Simulation and Green Solution for Urban Crime: Application of Large Language Model and Agent-Based Model

Urban environmental safety is critical for sustainable city development, yet the complex mechanisms underlying crime patterns pose significant research challenges. This study analyzes 12,893 criminal judgments (PCJs) from three Chinese megacities over three years using Large Language Models (LLM), developing a crime pattern simulation model that integrates Interpretable Machine Learning (IML) with Agent-Based Modeling (ABM). Then, three urban renewal scenarios were tested. The result shows that: (1) Socioeconomic and built-environment factors account for 33.3%–44.5% of outdoor crime spatial variation; (2) Nighttime lighting emerges as a key predictor, while canopy cover exhibits threshold effects on specific crime types; (3) Green regeneration scenarios demonstrate significantly lower crime rates. The proposed multidimensional framework provides actionable spatial planning insights and policy recommendations for safer cities.