Characterizing the Dynamic Hosting Capacity of Distribution Networks Integrated with Distributed GAI Data Center

Abstract

The widespread deployment of distributed Large Language Model (LLM) data centers across communities exposes distribution networks to significant stochastic power fluctuations driven by Generative Artificial Intelligence (GAI) training and inference tasks. Such random workload variations risk exceeding hosting capacity limits of Distribution Network (DN), while capacity expansion of critical equipment (e.g., transformers) to accommodate such volatility remains economically unjustifiable. To bridge this gap, the concept of Dynamic Hosting Capacity (DHC), enabling spatiotemporal characterization of distribution network constraints to optimize GAI data center integration. Detailed phase-specific power consumption models are established for LLM data centers, systematically quantifying their multidimensional spatiotemporal impact on DN safety margins through high-dimensional feasible region analysis. Furthermore, an efficient DHC characterization method based on the projection equivalence theory is introduced, which maps the high-dimensional feasible region onto the node-level dynamic hosting capacity indicators of the GAI interconnection nodes. Case studies based on a modified IEEE 33-bus feeder validate the effectiveness of the proposed method.