Risk Aware Reservoir Control For Safer Urban Traffic Networks

We present a risk-aware perimeter-style controller that couples safety and efficiency targets in large, heterogeneous urban traffic networks. The network is compressed into two interacting "reservoirs" whose dynamics follow the Generalized Bathtub Model, while accidents are described by a self-exciting (Hawkes) counting process whose intensity depends on vehicle exposure, speed dispersion between reservoirs and accident clustering. Accident occurrences feed back into operations through an analytically simple degradation factor that lowers speed and discharge capacity in proportion to the live accident load. A receding-horizon policy minimizes a mixed delay-safety objective that includes a variance penalty capturing risk aversion; the resulting open-loop problem is shown to possess a bang-bang optimum whose gates switch only at accident times. This structure enables an event-triggered MPC that only re-optimizes when new accidents occur, reducing on-line computation significantly. Parameters are calibrated using OpenStreetMap data for metropolitan Copenhagen to analyze traffic dynamics during morning peak commuter demand. Monte-Carlo simulations demonstrate delay savings of up to 30% and accident reductions of up to 35% relative to an uncontrolled baseline, with a transparent trade-off governed by a single risk parameter.