Meridia Insight Tech for Good Frontiers

The grid model that puts an AI data center on the other end of the wire

A data center is not one giant plug: this model traces AI demand through the UPS, cooling motors and transformers all the way to grid dynamics.

60%: the model assigns this share of initial data-center demand to IT, then traces it through UPS hardware into grid

A burst of AI computation can look, to the electric grid, less like a steady factory and more like a load that shifts under the system’s feet. This paper’s central contribution is a way to represent that shifting demand alongside the generators, inverters, transformers, cooling motors and backup power that shape how electricity actually moves. The ambition is practical: build a model detailed enough to capture fast interactions, but organized so simulations can solve them without getting stuck on abrupt switches in the equations.

That matters because data centers are no longer just large customers on a utility bill. Their servers, cooling equipment and power-conversion systems can respond on very different timescales, while the grid itself contains both slow-moving turbine-generator dynamics and fast electronic controls. A model that treats the facility as one fixed load can miss those interactions. Ali and Pandey instead describe a data center from its transmission connection down to the UPS that feeds its IT equipment, and set it inside a network model that includes conventional generators and grid-following inverters (Ali and Pandey, 2026).

But this is not a report that a particular grid became more reliable, or that a data center reduced its electricity use. The supplied document is supplementary modeling material for a separate paper on multirate power-system simulation. It lays out equations, components and initialization procedures; it does not provide a results section, a tested deployment, or numerical evidence that one simulation method outperforms another. Its value is the detailed, reproducible representation it offers for asking those questions.

The Science

Power grids are undergoing a change in what connects to them. A traditional generator converts mechanical energy into electricity through a spinning machine whose inertia gives the grid a kind of physical buffer. Solar, batteries and many modern loads connect through power electronics instead. Their converters can act quickly, but their behavior depends on control software and on what the surrounding network is doing. Meanwhile, AI data centers add concentrated loads whose demand can vary with computing work and whose cooling systems have their own electrical dynamics.

The paper assembles these pieces as two interacting layers. The slow layer is a positive-sequence transmission network: a simplified representation of three-phase alternating-current power systems that tracks the main voltage and phase behavior without following every individual waveform. Synchronous generators in that layer use the classical swing equation, which describes how a generator’s rotor angle and speed change when mechanical input and electrical output fall out of balance. A first-order turbine-governor model adjusts mechanical power in response to a setpoint and frequency deviation.

The fast layer represents grid-following inverters and data-center equipment with average models. “Average” means the model does not simulate every rapid on-off switching event inside the converter; it treats the power-electronic device as a controllable source whose behavior captures the slower control dynamics of interest. The two layers exchange voltage magnitude, phase angle, active power and reactive power at their connection buses. Active power does useful work; reactive power helps sustain the electric and magnetic fields needed by alternating-current equipment.

A grid-following inverter measures the grid and tries to synchronize with it. Its phase-locked loop, or PLL, estimates the grid’s phase angle by correcting a voltage error. A power controller turns requested active and reactive power into current targets. An inner current controller then adjusts the inverter’s voltage command to track those targets, while accounting for the coupling between the two axes of the rotating reference frame.

Figure 2: Grid-following inverter control block diagram.
Figure 2: Grid-following inverter control block diagram. Source: Muhammad Hamza Ali, Amritanshu Pandey

That controller architecture matters because an inverter is not simply a small generator. It follows an externally established grid reference rather than setting the voltage and frequency itself. The model includes the PLL, the power and current controllers, filters, and an output impedance. It assumes a constant DC-link voltage and omits switching dynamics—useful simplifications for system-level simulation, but not a complete description of semiconductor behavior or every control limit.

The data center is modeled as a chain of electrical equipment rather than a single demand number. Two step-down transformers connect the transmission bus to the facility’s low-voltage bus. That bus supplies three branches: the IT load, an induction-motor cooling load, and a constant-impedance load representing miscellaneous uses such as lighting and auxiliaries. The transformers’ leakage impedances are combined into one dynamic model, while the bus voltages are solved from the current balance among the branches.

Figure 3: Double-conversion UPS-based AI data center design.
Figure 3: Double-conversion UPS-based AI data center design. Source: Muhammad Hamza Ali, Amritanshu Pandey

The IT branch is supplied through a double-conversion UPS. In this arrangement, incoming AC is converted to DC by a rectifier, then converted back to AC by an output inverter that feeds the servers. A battery and bidirectional DC–DC converter connect to the DC side, allowing the modeled system to represent energy exchange with storage. The UPS output inverter is grid-forming on the server side: it establishes the local AC supply for the IT load rather than asking the transmission grid to set that output directly.

The authors drive IT demand with a measured GPU-power trace from an AI-inference dataset. They normalize the assigned trace so that the simulated data center begins at its chosen power-flow operating point, then interpolate between measurements. In other words, the workload signal changes the modeled power demand over time, while the power system responds to it. The excerpt provided does not include the trace itself, its duration, or quantitative simulation results.

The paper also addresses a numerical problem that can sound abstract but has real consequences for simulation. Power-system models often contain hard limits and other nonsmooth operations—rules with sharp corners where a tiny change in input can produce a discontinuous change in slope. Such behavior can frustrate nonlinear solvers, the algorithms used to find a consistent solution to many coupled equations. The authors replace nonsmooth operations with smooth surrogates so the discretized equations are continuously differentiable. That is a computational design choice, not proof that the physical devices themselves behave smoothly at every limit.

What They Found

The main result here is a model specification, not a measured grid outcome. The document gives equations for generator motion and governor response, inverter synchronization and current control, transformer dynamics, load behavior, cooling-motor fluxes, and UPS elements. It also describes how to initialize these devices at a power-flow operating point—the steady-state condition where the network’s electrical supply and demand balance—so a simulation does not begin with an artificial transient simply because its components start out inconsistent.

That full-system view is the distinctive feature. A computing workload can change the power drawn by servers; the UPS can mediate what the servers see; the transformer and cooling branches contribute their own demand; and the grid-side inverter or generator controls respond to resulting electrical conditions. Instead of treating those as disconnected engineering topics, the model makes their interfaces explicit. Its authors describe the interaction between slow and fast subsystems through the voltage, angle and power quantities at their coupling buses.

The facility assumptions also give a sense of scale, though they should not be mistaken for measured results. The model sets the IT load at 60% of the initial low-voltage-bus active-power demand and the miscellaneous impedance load at 10%. Separately, it cites cooling as typically about 30% of a standalone data center’s total power use. Those figures are not a reported measurement of one facility, and the cooling share is presented as a typical estimate rather than a fitted result from the GPU trace.

Load shares used or cited in the data-center model

Reported model allocation and a separately cited typical estimate. The cooling figure is approximate and should not be read as a measured share in the modeled facility.

Load shares used or cited in the data-center model
LabelValue
IT load at initial LV demand60 % of initial LV active-power demand
Miscellaneous load10 % of initial LV active-power demand
Cooling, typical estimate cited30 % of initial LV active-power demand

The architecture itself is concrete: two cascaded transformers, three low-voltage branches, and UPS components including a rectifier, output inverter, battery and bidirectional DC–DC converter. Those counts describe what is represented in the model, not the capacity or prevalence of data centers in the world.

What the modeled data center contains

Counts of modeled structural elements; these are architecture counts, not power ratings.

What the modeled data center contains
LabelValue
Cascaded step-down transformers2 transformers
Parallel branches on LV bus3 transformers
Named UPS elements4 transformers

The work’s other important finding is methodological: initialization is specified component by component. For example, transformer current starts from the power-flow active and reactive demand so that the initial current derivative is zero. That helps align the dynamic model with the pre-existing network operating point. Such care is essential when comparing a simulated response with a real disturbance: otherwise, a model can appear to respond dramatically just because it was improperly started.

What the document does not establish is equally important. It does not show how much a GPU workload trace moves total grid demand, whether UPS batteries smooth that demand, whether cooling motors amplify a voltage disturbance, or how a network performs under a particular fault. Nor does it report accuracy against field measurements. Those are natural questions for the companion simulation study, but the supplied supplement alone cannot answer them.

Why This Changes Things

For grid planners, the choice of load model can alter the question they are able to ask. A fixed megawatt value says how much electricity a facility uses at one instant; it says little about how its demand changes, which equipment controls it, or how it behaves when voltage dips. This model creates a framework for testing those dynamics—especially the interplay between power electronics and conventional grid machinery—rather than assuming a data center is electrically inert.

That could matter as utilities connect large, fast-growing computing campuses. AI workloads are not identical to the electricity demand of a city or a steel mill. The server load can shift with computation; UPS systems can separate server supply from disturbances on the AC side; and cooling motors respond to voltage and frequency in ways that a simple constant-power block cannot capture. A detailed model can help researchers test whether these features damp, transmit or complicate disturbances. The paper does not claim any one outcome, but provides a more physically explicit starting point for finding out.

It also puts the UPS in a broader role. Backup power is usually discussed as insurance for the servers: keep electricity flowing when the grid fails. Here, the UPS is also a modeled interface between computation and the grid. Its rectifier, DC bus, battery converter and grid-forming output inverter determine how demand reaches the facility’s AC connection. That representation can support future studies of how storage controls or workload timing might be coordinated with grid needs, while recognizing that actual UPS operation depends on design, settings, battery limits and reliability requirements.

There is a related shift in how power systems are simulated. A transmission grid’s generators evolve relatively slowly, while converter controls can act much faster. Solving every component at the fastest timescale can become computationally expensive; using one coarse timestep can miss important control behavior. This model is built for a research context focused on adaptive multirate waveform relaxation—a technique intended to let different parts of a simulation advance at suitable rates while exchanging boundary information. The supplement supplies models for that exercise; it does not itself quantify the speedup or accuracy of the technique.

Smoothing sharp operations illustrates the tradeoff between mathematical tractability and physical detail. Smooth equations are easier for many nonlinear solvers to handle, but a surrogate must be chosen carefully so that it approximates the actual control limit or device behavior in the range that matters. The paper’s supplementary excerpt says the replacements make the discretized model continuously differentiable; it does not provide enough detail here to evaluate the size of any approximation error. Reproducibility will depend on publishing the exact surrogates, parameters and validation results alongside the equations.

What’s Next

The next step is validation. Researchers would need to compare the modeled response with equipment data or controlled tests: transformer and motor behavior, UPS input and output power, inverter controller response, and the effect of real workload traces on a facility’s total demand. Sensitivity tests could show which assumptions—constant DC-link voltage, averaged converters, load composition, cooling-motor parameters—most affect the predicted grid response.

The control questions are open, too. A UPS battery might provide flexibility, but using it for grid support could compete with keeping reserve for an outage. A workload could perhaps be scheduled to reduce a sharp demand ramp, but computing deadlines and service quality constrain that flexibility. A model linking measured GPU demand to the actual AC-side response is a useful foundation for exploring those possibilities; it is not evidence that they are already practical or beneficial.

There are also limits in the scope of the excerpt. It focuses on a positive-sequence network and average converter models, rather than unbalanced faults, switching-level behavior, or every operational protection mechanism. Its cooling representation uses an induction motor, while real data centers can have varied cooling architectures and control strategies. Results from this model will therefore need to be interpreted in light of the facility and network it represents.

Still, the contribution is valuable precisely because it makes the electrical boundary around an AI data center less of a black box. By connecting workload-driven IT demand to UPS equipment, cooling, transformers, inverters and synchronous generators, Ali and Pandey offer researchers a system they can interrogate and improve. As computation becomes a more consequential part of electricity demand, models that connect server behavior to grid dynamics can help turn vague concern into testable engineering questions—and, eventually, better decisions about how large new loads join the grid.

Comments (0)

No comments yet. Be the first to share your thoughts.