Large Language Model-Driven Traffic Signal Optimization for Reducing Energy Consumption and Urban Pollution

Urban traffic congestion directly contributes to excessive energy consumption and urban air pollution, requiring adaptive traffic signal control strategies that incorporate sustainability objectives alongside mobility performance. This study proposes a Large Language Model (LLM) driven traffic signal optimization framework that transforms detailed intersection-level traffic states into structured natural-language prompts, enabling the LLM to reason over congestion patterns, queue asymmetry, phase history, and estimated energy emission impacts. Unlike reinforcement learning (RL) based controllers, the LLM requires no task-specific training and operates in a zero-shot manner through carefully designed structured prompts that encode traffic states, phase history, and control constraints, enabling interpretable and context-aware decision-making. The framework is evaluated using both single-intersection and multi-intersection scenarios in the CityFlow simulator. To quantify environmental impact, energy consumption and emissions are estimated using a trajectory-based approximation model that applies aggregated coefficients for idling, cruising, and stop-and-go events. Experimental results demonstrate that the proposed LLM-based controller achieves substantial improvements in sustainability and mobility metrics. GPT-4 reduces average per-vehicle energy consumption to 7.94 MJ, representing a 29% improvement over fixed-time control and a 19.7% decrease in total network energy usage. GPT-4.1-mini achieves the shortest average travel time at 278.03 s, outperforming state-of-the-art RL baselines while maintaining competitive energy efficiency. The LLM also reduces idle time by 26.2%, compared to the fixed-time baseline, contributing directly to lower stop-and-go emissions. We adopt an API-based LLM in our experiments to enable a reproducible assessment of runtime feasibility for LLM-driven traffic signal control. With a 30 s decision interval per phase, the end-to-end API response time remains compatible with real-time actuation; moreover, future self-hosted/on-premises deployment is expected to further reduce latency without altering the control interval. We also discuss practical cost considerations for continuous operation. Despite these promising results, LLM-based control can be sensitive to prompt formulation and may occasionally yield hallucinated or unsuitable actions. Accordingly, real-world deployment in safety-critical infrastructure should incorporate explicit safety constraints, runtime monitoring, output validation, and a deterministic fallback controller. Overall, the proposed framework supports multi-objective optimization by jointly balancing mobility (e.g., delay and throughput) and sustainability (e.g., energy use and emissions) through a unified reward-guided decision policy, while providing more interpretable decision rationales under appropriate safety guardrails.