Minute-scale incident mitigation in urban networks via threshold allocation and AI planning a three-stage corridor framework

Urban traffic agencies must convert optimisation outputs into concrete control within minutes after an incident. We present a three-stage framework that bridges this optimisation-to-action gap. Stage-H harvests a compact set of parallel source–sink corridors with bounded overlap, yielding interpretable route supports. Stage-M allocates inflow via a threshold (min–max) rule by solving for a unique water-filling level z* so that the sum of capped flows meets the target demand; the allocation respects per-corridor capacity derived from storage and signal feasibility and can be computed in near-linear time. Stage-P compiles target flows into executable actions—signal-gating schedules and detour plans—using domain-independent planning with explicit timeouts and fallbacks. We evaluate on a 20×20 grid and a city-clip network in SUMO, against dynamic re-routing, static gating and perimeter/MFD-style baselines, and under stressors including measurement latency and elastic driver compliance. Across networks and scenarios, the pipeline improves clearance and tail delay while maintaining corridor-level feasibility, and planner wall-times remain compatible with minute-scale control windows. Components are modular and the full artefact package (data, code and figure scripts) enables end-to-end reproducibility.