LLM and Deep Learning in the Loop of Disturbed Traffic Control

Traffic signal control increasingly faces disturbed operating conditions such as incidents, abrupt demand surges, sensing degradation, and abnormal driving patterns. Under these nonstationary regimes, classical fixed-time and actuated strategies may exhibit slow recovery, while purely data-driven controllers can be brittle when disturbance characteristics shift. This paper proposes an LLM-in-the-loop architecture for disturbed traffic signal control that integrates (i) deep learning for disturbance detection and short-horizon traffic forecasting, (ii) a disturbance-aware candidate generation and scoring layer (template/retrieval-based), and (iii) a constrained large language model (LLM) that selects or minimally repairs signal plans only within constraint-screened action templates. A deterministic validator enforces safety and operational constraints, including minimum/maximum greens, cycle feasibility, and clearance rules, by checking action feasibility before execution. The method is formulated as constrained decision making under uncertainty, where disturbance estimates and predictive confidence shape both retrieval/scoring and LLM supervision. The originally reported SUMO evaluation considered multiple disturbance categories, including capacity drops, demand shocks, and sensing dropouts as well as reported network delay, queue spillback, recovery time, and switching stability. Within the originally reported SUMO scenarios, descriptive results suggest that among the selected baselines, the proposed DL + LLM framework reported lower mean values of delay, spillback frequency, and recovery time than the fixed-time, actuated, and retrieval-only baselines. The reported validator-detected action-feasibility violations were zero; this result concerns timing-action feasibility rather than an absence of traffic-state risks such as spillback.