Aiming at the problem of motor vehicle energy consumption analysis at urban road intersections, this study establishes a large language model (LLM) prompt framework based on prompt engineering and chain of thought (CoT). Using the intersection energy consumption dataset of the central urban area of Cangzhou City provided by Amap, the motor vehicle energy consumption level at intersections is selected as the core indicator to characterize the energy consumption status of intersections. The analysis effects of four large language models (GPT-4o, GPT-3.5, Claude, LLaMA-2) are evaluated in scenarios of zero-shot learning, few-shot learning, and prompt engineering enhanced by chain of thought. Results show that the established LLM prompt framework enables task completion across four scenarios, with GPT-4o outperforming other models, few-shot learning proving more effective than zero-shot learning, and the chain of thought technique significantly boosting LLM performance in intersection energy consumption analysis. The results indicate that the established large language model prompting framework successfully completes analysis tasks across four scenarios, with GPT-4o demonstrating optimal performance—achieving the highest F1 score of 0.5461 in zero-shot learning with chain-of-thought prompting. In basic prompting engineering, few-shot learning outperforms zero-shot learning, significantly boosting the framework’s training efficacy with F1 score improvements of 3% to 61%. Notably, chain-of-thought techniques enhance model performance in intersection energy consumption analysis, yielding average F1 score gains of 12% to 30%.