Multi-Tenant Edge AI as a Virtual Power Plant via Online Auction-Driven Energy Scheduling

Abstract

We innovatively propose that the modern edge computing infrastructure equipped with battery energy storage can actively operate as a Virtual Power Plant (VPP) to support grid stability through real-time energy dispatch. However, the edge infrastructure does not directly control the workload of the edge AI services that rent its resources, making it challenging to align energy flexibility with grid needs. To bridge this gap, we introduce an auction-driven online framework that incentivizes edge AI services to dynamically adapt their inference models in exchange for payment, thereby freeing battery capacity for grid support. Our approach tackles a novel NP-hard long-term social-cost minimization problem that jointly optimizes energy scheduling, battery degradation, inference accuracy, and model-switching overhead over time. Addressing the fundamental challenges of battery state transitions, complex switching cost across auctions, and economic guarantees for each single auction, we design a polynomial-time algorithmic framework that integrates switching-cost-aware control, online learning, randomized rounding, and a truthful payment mechanism. Theoretically, we rigorously prove that our online algorithms achieve an asymptotic competitive ratio, sub-linear regret and fit, and truthfulness and individual rationality. Extensive practical experiments using real-world traces demonstrate that our approach not only achieves a competitive ratio of 1.63 against the offline optimal solution while yielding verifiable sub-linear regret and fit curve, but also reduces social cost substantially compared to heuristics and state-of-the-arts and scales efficiently under dynamic workloads.