⚡ electrical engineering

Grid Integration of Gigawatt-Scale AI Data Centers under Connect-and-Manage

This paper proposes a hierarchical coordination framework for gigawatt-scale AI data centers under "connect-and-manage" interconnection rules, which integrates differentiated workload flexibility and on-site storage to significantly reduce grid curtailment while maintaining critical training workloads despite opaque transmission system operator acceptance mechanisms.

Original authors: Xin Lu, Qianwen Xu

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xin Lu, Qianwen Xu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine a massive, super-fast AI factory (a Gigawatt-scale AI Data Center) that needs to plug into the national power grid to do its work. In the past, building a connection to the grid was like buying a permanent, unlimited highway lane: you paid for the construction, and you could drive as much as you wanted, whenever you wanted.

But the world is changing. These AI factories are growing so fast that the power grid can't be upgraded quickly enough to handle them all. So, a new rule called "Connect-and-Manage" is emerging.

Think of this new rule like getting a "pay-as-you-go" data plan for electricity instead of an unlimited one. You can plug in, but the power company (the Grid Operator) might occasionally tell you, "Sorry, the road is too crowded right now; you need to slow down or stop for a bit."

This paper proposes a smart system to help these AI factories handle those "slow down" orders without crashing their work. Here is how it works, broken down into simple parts:

1. The "Request-and-Reply" Game

Instead of just guessing how much power they need, the AI factory and the power grid play a daily game of Request and Reply:

The Request: The AI factory asks, "Can I have 1,000 megawatts of power right now?"
The Reply: The Grid Operator looks at the traffic on the power lines. If the lines are clear, they say, "Yes, take it all!" If the lines are jammed, they say, "No, you can only have 900 megawatts. We are cutting off 100 megawatts to keep the grid safe."
The Catch: The AI factory doesn't know exactly how the Grid Operator makes these decisions. It's like trying to guess the traffic light pattern without seeing the sensors. The AI has to learn by trial and error.

2. The AI Factory's "Smart Brain" (Three Layers)

To handle this uncertainty, the paper designs a three-layer "brain" for the AI factory:

Layer 1: The Strategist (Planning)
This is the long-term planner. It looks at the weather, the time of day, and past grid behavior to guess, "Hey, it's usually busy at 2 PM, so maybe I should ask for less power then." It uses Reinforcement Learning (a type of AI that learns from mistakes) to get better at predicting when the grid will say "no."
Layer 2: The Gatekeeper (Grid Operator)
This is the power company. They check if the request is safe. If it's not, they cut the power (curtailment) and tell the factory the new limit.
Layer 3: The Tactician (Execution)
This is the on-the-ground manager. Once the factory gets the "cut" notice, this layer instantly rearranges the work. It decides which computers can slow down and which must keep running at full speed.

3. The "Three Types of Workers" (Workloads)

The paper realizes that not all AI work is the same. It divides the factory's work into three groups, like a restaurant kitchen:

The "Frontier" Chefs (Frontier Training): These are working on a massive, complex recipe (like training a new super-AI) that takes days or weeks. If they stop, the whole recipe is ruined. They are the VIPs. The system protects them at all costs.
The "Batch" Prep Cooks (Batch Training): These are doing smaller tasks, like chopping vegetables or testing small recipes. If they stop, it's annoying, but they can just start again later. They are the flexible ones. When the grid gets crowded, these are the first to be asked to slow down.
The "Inference" Waiters (Inference): These are serving customers right now (answering questions). They can't really stop; they just serve as many customers as they can with the power they have.

4. The "Battery Buffer" (The Emergency Kit)

The factory also has a giant on-site battery. Think of this as a backpack full of extra snacks.

When the grid cuts power: The factory opens the backpack and uses the stored energy to keep the "Frontier Chefs" working while the "Batch Prep Cooks" take a break.
When the grid is quiet: The factory fills the backpack back up, waiting for the next emergency.

5. The Results: What Happened?

The researchers tested this system on a simulated power grid (based on the Australian market and a standard electrical test system). Here is what they found:

Less Cutting: Before this smart system, the AI factory had its power cut about 9.1% of the time. With the new system, cuts dropped to just 2.8%.
Protecting the VIPs: Even when cuts happened, the system managed to keep the "Frontier" work running at 98.1% of its target.
Smart Sacrifices: The system learned to sacrifice the "Batch" work (the flexible prep cooks) to save the "Frontier" work. The "Batch" throughput swung wildly up and down, acting like a shock absorber for the grid.
Battery Magic: The battery didn't just sit there; it actively discharged during cuts and deferred charging when the grid was stressed, acting as a buffer to smooth out the bumps.

Summary

In short, this paper teaches a massive AI factory how to dance with the power grid. Instead of fighting the grid when it gets crowded, the factory learns to ask politely, predict the traffic, and sacrifice the less important tasks to keep the most important work running. It uses a smart battery as a shock absorber and a learning AI to figure out the best way to ask for power without getting rejected.

Technical Summary: Grid Integration of Gigawatt-Scale AI Data Centers under Connect-and-Manage

Problem Statement
The rapid deployment of gigawatt-scale artificial intelligence data centers (AIDCs) creates a structural misalignment with the multi-year timeline of transmission grid reinforcement. To bridge this gap, emerging "connect-and-manage" frameworks allow AIDCs to interconnect without prior grid upgrades, subjecting them to real-time curtailment during grid stress. However, existing literature lacks a formal representation of the bilateral interaction between the AIDC and the Transmission System Operator (TSO) under this regime. Current models either treat data centers as passive demand-response resources or assume full information symmetry, failing to capture the "connect-and-manage" reality where the TSO's acceptance decision is a black box to the AIDC. Furthermore, existing physical models often homogenize AIDC workloads, ignoring the distinct flexibility of frontier training, batch training, and inference, and neglecting the role of on-site battery energy storage systems (BESS) as a buffer.

Methodology
The paper proposes a formalized framework comprising three core components:

Interaction Protocol and Information Boundary:
The authors formalize the AIDC–TSO coordination as a sequential request–acceptance protocol. At each time step $t$ , the AIDC submits a power request $P_{req}(t)$ . The TSO evaluates this against network constraints (modeled via DC power flow with generator limits and a budget-constrained demand uncertainty set) and returns an accepted exchange $P_{acc}(t) = P_{req}(t) - \kappa(t)$ , where $\kappa(t)$ is the curtailment variable. Crucially, a strict information boundary is established: the AIDC observes the accepted power and public signals but lacks visibility into the TSO's internal network topology, generator costs, or optimization logic. This renders the TSO's acceptance mapping a black box, precluding standard bilevel optimization.
Heterogeneous Physical Modeling:
- AIDC Side: The facility is decomposed into three workload subclasses: Frontier training (large-scale foundation models requiring continuity), Batch training (smaller, interruptible tasks), and Inference (exogenous demand tracking). An on-site BESS is integrated to buffer the gap between grid-accepted power and internal IT load.
- Grid Side: The TSO's feasibility is modeled using a robust optimization approach with a budget uncertainty set to account for demand deviations, ensuring security under worst-case scenarios within a defined budget.
Hierarchical Decision Architecture:
To navigate the information asymmetry, a three-layer architecture is developed:
- Planning Layer (Learning-Based): A Soft Actor-Critic (SAC) reinforcement learning agent learns a capacity-aware policy. It observes historical interaction outcomes (curtailment patterns) and system signals to generate power requests that anticipate TSO constraints.
- TSO Layer (Robust Acceptance): The TSO solves a security-constrained robust optimization problem to determine the accepted power, minimizing curtailment and generation costs while adhering to network limits.
- Execution Layer (Optimization): A single-step mixed-integer linear program (MILP) translates the planning targets into feasible internal operations. It enforces exact power balance, manages BESS dispatch, and reallocates the reduced power budget across clusters, prioritizing Frontier training over Batch training based on asymmetric penalty weights.

Key Contributions

Formalization of Connect-and-Manage: The paper establishes a sequential request–acceptance protocol with an explicit curtailment variable and a defined information boundary, providing the first formal representation of this emerging operational regime.
Differentiated Physical Modeling: It introduces a granular AIDC model that distinguishes between Frontier, Batch, and Inference workloads with distinct operational constraints and penalty structures, while integrating on-site BESS as a dynamic power buffer.
Hierarchical Decision Framework: It proposes a novel three-layer architecture that combines reinforcement learning (to handle the black-box nature of the TSO) with robust optimization (for grid security) and execution optimization (for internal feasibility), effectively solving the decision problem under information asymmetry.

Results
Case studies utilizing a gigawatt-scale AIDC on the IEEE 39-bus system with Australian market data demonstrate the framework's efficacy:

Curtailment Reduction: The proposed learning-based approach reduced curtailment frequency from 9.1% (under a static buffer strategy) to 2.8%.
Workload Preservation: Despite grid stress, the framework preserved 98.1% of the Frontier training workload, significantly outperforming heuristic and other deep reinforcement learning baselines (DDPG, TD3).
Resource Differentiation: The results confirm that Batch training acts as the primary grid-elastic resource, exhibiting the largest throughput swings during peak demand, while Frontier training is protected, and Inference remains demand-driven.
BESS Functionality: The on-site BESS serves as a critical buffer, providing active discharge during severe curtailment events and deferring charging to absorb residual constraints, thereby minimizing throughput loss.

Significance and Claims
The paper claims that its framework provides a necessary methodology for AIDC decision-making under connect-and-manage, a setting where standard bilevel and stochastic optimization approaches fail due to information asymmetry. By formalizing the interaction protocol and developing a learning-based planning layer coupled with robust grid modeling, the work demonstrates that gigawatt-scale AIDCs can operate flexibly and reliably without requiring prior grid upgrades. The study highlights that the combination of heterogeneous workload management and on-site storage allows AIDCs to act as effective grid-elastic resources, preserving critical AI workloads while accommodating transmission constraints. The authors note limitations regarding the 15-minute resolution and the exclusion of supply-side renewable uncertainty, suggesting future work in multi-AIDC coordination and AC power flow.