ARTFEED — Contemporary Art Intelligence

HetGPS: Physics-Anchored Safety for EV Charging Networks

ai-technology · 2026-08-04

A novel framework named HetGPS has been developed to enhance safety measures for extensive groups of network-connected agents, particularly in the context of electric vehicle (EV) charging. This framework integrates learned graph risk with physics-based corrections by differentiating the magnitude of interventions from their directional guidance. An action-conditioned graph residual model manages state-dependent intervention authority, while a physics model dictates the intervention's direction. For EV charging, HetGPS works alongside a parameter-shared heterogeneous graph soft actor-critic policy, facilitating topology-aware coordination that remains unaffected by fleet size. In tests across five nested distribution networks involving 200–3,218 EVs over 100 evaluation days, the Adaptive Authority approach decreased bus-step voltage violations from 3.93–7.74% to 0.52–3.44%, achieving a departure success rate of 99.06–100%. The research can be found on arXiv under identifier 2608.00679.

Key facts

  • HetGPS is a hybrid graph-control framework for safety interventions.
  • It separates intervention magnitude from corrective direction.
  • An action-conditioned graph residual model schedules intervention authority.
  • A physics model determines the corrective direction.
  • The framework is applied to EV charging with a heterogeneous graph soft actor-critic policy.
  • Evaluations involved five nested distribution networks with 200–3,218 EVs.
  • Voltage violations reduced from 3.93–7.74% to 0.52–3.44%.
  • Departure success maintained at 99.06–100%.
  • The paper is available on arXiv (2608.00679).

Entities

Institutions

  • arXiv

Sources