⚡ electrical engineering

Battery-Assisted Operation of Hyperscale AI Data Centers under Connect-and-Manage Interconnection Practices

This paper proposes a battery-assisted operational framework that integrates on-site energy storage with a continuity-aware energy-computation model and a two-stage control strategy to enable hyperscale AI data centers to reconcile rapid internal workload and cooling dynamics with real-time, time-varying grid interconnection limits, thereby enhancing both workload commitment and delivery robustness under transmission congestion.

Original authors: Xin Lu, Jing Qiu, Jiafeng Lin, Sihai An, Mingyang Sun, Junhua Zhao

Published 2026-05-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xin Lu, Jing Qiu, Jiafeng Lin, Sihai An, Mingyang Sun, Junhua Zhao

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

The Big Picture: The "Connect-and-Manage" Problem

Imagine a massive, high-tech factory (an AI Data Center) that needs a huge amount of electricity to run its supercomputers. This factory is so big it's like a city block of power usage.

In the old days, before building this factory, the power company would say, "We need to check our roads (power lines) first. If they can't handle the traffic, we can't let you build." This is slow and expensive.

Now, a new rule called "Connect-and-Manage" is emerging. It's like saying, "Okay, you can build your factory and connect to the grid right away! But, if the roads get too crowded, we will put up a temporary speed limit or a gate that only lets a certain amount of traffic through at specific times."

The Problem:
AI training is like a marathon runner who can't stop running in the middle of a race without losing their place. If the power company suddenly closes the gate (lowers the power limit), the factory has to stop. But stopping AI training is bad because it loses progress. Also, the factory generates a lot of heat, so it needs air conditioning, which also uses power.

The factory is stuck between two conflicting worlds:

Inside: It needs steady, continuous power to keep the computers running and the heat down.
Outside: The power grid is unpredictable and sometimes says, "You can only use 50% of your power right now."

The Solution: The "Battery Buffer"

The paper proposes a solution: Put a giant battery (BESS) right next to the factory.

Think of the battery as a shock absorber or a water tank between the factory and the power grid.

When the grid is tight: The grid says, "You can only take 100 units of power." The factory actually needs 200. The battery steps in, says, "I'll give you the other 100," so the factory keeps running without stopping.
When the grid is loose: The grid says, "You can take 300 units." The factory takes what it needs, and the battery fills up with the extra power (or sells it back to the grid for profit).

This battery acts as a physical buffer, smoothing out the bumps so the factory doesn't have to panic every time the grid changes its mind.

How They Plan It: The Two-Step Dance

The researchers created a smart two-step plan to make this work:

Step 1: The Day-Ahead Promise (The Forecast)
The day before, the factory manager looks at the weather, electricity prices, and historical traffic patterns. They use a computer model to guess how strict the power grid will be tomorrow.

Based on these guesses, they make a promise: "We promise to finish this much AI work tomorrow."
They calculate the safest amount of work they can promise, assuming the grid might get tight. They don't promise too much, or they might fail.

Step 2: The Real-Time Dance (The Execution)
On the day of, the grid sends real-time updates: "Okay, right now, the limit is tight," or "Now, the limit is wide open."

The factory uses a smart controller (like a self-driving car) that looks at the next few hours.
If the grid gets tight, the controller instantly tells the battery: "Discharge now to keep the computers running!"
If the grid is open and electricity is cheap, the controller tells the battery: "Charge up!" or "Turn on the air conditioning early to cool things down while it's cheap."

What They Found (The Results)

The researchers tested this idea using a simulated power grid and real data from Australia. Here is what happened:

More Work Gets Done: With the battery, the factory could promise to do 38% more work the day before compared to not having a battery. Without the battery, the factory had to stop working whenever the grid got crowded.
Fewer Stoppages: Without the battery, the factory had to shut down on 17 out of 31 days. With the battery, it only had to stop on 5 days. The battery acts as a safety net.
The Battery Changes Roles:
- When the grid is very crowded: The battery acts like a life raft. It just does whatever it takes to keep the computers running, even if it costs a bit more. Its only job is survival.
- When the grid is relaxed: The battery acts like a smart investor. It waits for cheap electricity to charge up and sells power when prices are high, saving the factory money.
Size Matters: If the battery is too small, it runs out of juice during a long power outage, and the computers still have to stop. A bigger battery allows the factory to "pre-cool" the building (turn on AC early) and then survive on battery power alone during the crunch time.
The "Checkpoint" Rule: AI training can only be safely paused at specific moments (like saving a game). If these "save points" are far apart, the battery has to work harder to keep the machine running until the next save point. If save points are frequent, the factory can just pause and wait, needing less battery help.

The Bottom Line

This paper shows that for massive AI data centers to connect to the power grid without causing blackouts or getting shut down, they need a giant battery on-site. This battery acts as a translator and a shock absorber, allowing the factory to keep running its complex AI tasks even when the power grid is having a bad day. It turns a volatile, unpredictable power connection into a reliable, smooth operation.

Technical Summary: Battery-Assisted Operation of Hyperscale AI Data Centers under Connect-and-Manage Interconnection Practices

1. Problem Statement

The rapid deployment of hyperscale Artificial Intelligence Data Centers (AIDCs) presents a significant challenge to modern power systems. A single AIDC can demand hundreds of megawatts, yet the planning cycles for transmission infrastructure often lag behind AIDC deployment. Consequently, grid operators are shifting from traditional "review-first, connect-later" paradigms to "connect-and-manage" practices. Under this framework, new mega-loads are permitted to connect but are subject to real-time, time-varying admissible power exchange limits at the Point of Common Coupling (PCC) to ensure transmission reliability.

This creates a fundamental conflict for AIDCs:

Internal Dynamics: AIDCs hosting large-scale AI training workloads are inherently stateful. Interruptions are only permissible at discrete checkpoint moments. Furthermore, IT computing loads are tightly coupled with cooling systems, creating fast internal thermal and power dynamics.
External Constraints: The grid imposes fluctuating power exchange limits based on network congestion.
The Mismatch: Directly mapping internal AIDC dynamics to external grid limits is problematic. If the grid tightens limits, a direct load curtailment strategy risks frequent, unsafe computational interruptions and service delivery failures. Existing load flexibility models often assume loads can be switched arbitrarily, an assumption that does not hold for checkpoint-constrained AI training.

2. Methodology

The paper proposes a battery-assisted operational framework that utilizes on-site Battery Energy Storage Systems (BESS) as a physical buffering interface to decouple fast internal AIDC dynamics from time-varying grid constraints.

A. Energy-Computation Modeling
The authors develop a joint model capturing three coupled layers:

Computing Layer: AIDC workloads are modeled as stateful, batch-oriented AI training. The model enforces checkpoint-aware continuity, where the computing cluster can only be shut down at specific time slots ( $\delta_{ckpt}$ ). Throughput is regulated via software-defined power limits, exhibiting a strictly convex relationship between electrical power and computational throughput.
Thermal Layer: A first-order RC thermal circuit model describes the data hall temperature dynamics. Cooling power is coupled with IT heat generation, allowing for thermal inertia exploitation (e.g., pre-cooling) to shift cooling loads.
Grid Interface: The PCC power exchange is constrained by time-varying upper and lower limits ( $P_{grid}^{min}(t), P_{grid}^{max}(t)$ ) and ramp-rate requirements imposed by the Transmission System Operator (TSO).

B. Two-Stage Decision Framework
To address information asymmetry between day-ahead planning and real-time operation, a two-stage logic is formulated:

Stage I: Day-Ahead (DA) Workload Commitment:
- Using historical data and publicly observable signals (prices, demand, temperature), a conditional generative model (CTSGAN) predicts a set of admissible PCC power limit trajectories.
- A scenario-based optimization determines a single scalar workload commitment ( $W_{DA}^*$ ) that is credibly deliverable across all retained scenarios (filtered for a 90% confidence level). This stage maximizes the committed workload while penalizing deviations from energy-efficient operating points.
Stage II: Real-Time (RT) Delivery Assurance:
- A receding-horizon controller operates at 15-minute intervals.
- It receives realized grid limits and prices, then optimizes the dispatch of IT throughput, cooling, and BESS charging/discharging.
- The objective is to minimize operational costs (energy + battery degradation) while strictly adhering to grid constraints and ensuring the delivery of the DA-committed workload. A large penalty is applied to any under-delivery slack.

3. Key Contributions

Novel Modeling Approach: The paper presents an energy-computation model that jointly characterizes stateful, checkpoint-constrained AI training workloads and tightly coupled IT-cooling thermal dynamics, providing a coherent representation of AIDC operational continuity under power constraints.
Physical Buffering Framework: It proposes a grid-aware framework where on-site BESS acts as a physical interface, enabling time-scale decoupling between fast internal computing-cooling dynamics and external transmission constraints.
Two-Stage Decision Logic: A coordinated mechanism is developed to align day-ahead workload commitments with real-time grid feasibility, ensuring commitments remain credible under uncertain transmission congestion while maintaining strict grid compliance.

4. Results

Case studies were conducted on the IEEE 39-bus system using real data from the Australian Energy Market (AEMO) for January 2026.

Enhanced Commitment Capability: Under transmission congestion, the BESS-assisted framework significantly increased the credible day-ahead workload commitment. In a representative congested scenario, the AIDC with BESS sustained ~175 MW operation during periods where the grid limit collapsed to 0 MW, whereas the non-BESS case faced 7.75 hours of complete service interruption.
Delivery Robustness: Over a 31-day simulation, the BESS-assisted system reduced workload under-delivery from 17 days (no BESS) to 5 days. The average under-delivery was reduced by an order of magnitude, demonstrating improved robustness against inter-day uncertainty.
Regime-Dependent BESS Role:
- Feasibility-Oriented: Under tight transmission constraints, the BESS primarily acts as a continuity bridge, discharging to sustain IT operations when grid imports are restricted.
- Economy-Driven: As transmission constraints relax, the BESS role transitions to economic flexibility, exploiting price signals (e.g., exporting power during high-price periods) to reduce overall energy costs without compromising delivery.
Capacity and Checkpoint Sensitivity:
- Sufficient BESS energy capacity is critical for enabling "pre-cooling" strategies, where thermal inertia allows cooling loads to be shifted, prioritizing power for IT computation during grid congestion.
- The frequency of checkpoints dictates BESS utilization; sparse checkpointing forces the BESS to provide sustained support to bridge gaps between admissible shutdown windows, while frequent checkpointing allows for more flexible load shedding with less reliance on storage.

5. Significance and Claims

The paper claims that the proposed framework offers a physically grounded and operationally tractable approach for integrating large-scale AIDCs into constrained transmission networks. By explicitly positioning BESS as a physical interface, the framework resolves the mismatch between internal operational flexibility and grid security requirements.

The findings suggest that under "connect-and-manage" practices, BESS is not merely an economic asset but a critical enabler of operational continuity. It allows AIDCs to make credible day-ahead commitments despite uncertain grid conditions and ensures reliable delivery in real time. The study highlights that the value of BESS is regime-dependent: it is essential for feasibility when transmission limits are binding and shifts to economic optimization when constraints are relaxed. This provides actionable insights for the co-design of future power systems and computation-intensive infrastructures.