D1.3 · Understand AI load profiles and timescales
A connection application, a battery energy estimate and a power-quality study ask different questions of the same load. This lesson explains how to choose the right measurement boundary and time resolution before drawing conclusions. You will use a deliberately simplified load cycle to see what an average hides, then turn the remaining uncertainty into a practical data request.
What should you be able to judge after this lesson?
- Distinguish an interval-average demand, an observed peak, a ramp rate and an energy requirement.
- Explain which claims a slow meter can support and which need faster, synchronised measurements.
- Prepare a load-data request that protects service requirements and records evidence quality.
1. Define where the load is measured before comparing numbers
A rack reading describes equipment behind that rack meter. An IT distribution-board reading includes a different collection of equipment, and the utility point of common coupling also sees cooling, electrical losses and any on-site generation or battery action. A spike at one boundary need not have the same amplitude or timing at another. Keep the meter location, sign convention, units and timestamp beside every series.
This lesson uses total facility AC demand measured upstream of a proposed battery connection. It includes facility auxiliaries but excludes battery losses. Grid import equals this demand minus battery discharge. We do not multiply the figures by annual PUE: an annual energy ratio does not determine the coincident power during a short event.
2. Match the time resolution to the decision
For a commissioning plan, monthly or phase-based demand helps establish when capacity is needed. For a demand charge or an energy budget, preserve the tariff interval and calculate energy over that interval. For a battery dispatch study, use a time series that resolves the events the controller is expected to manage. These datasets can support one another, but one is not a substitute for all the others.
Voltage disturbances, waveform distortion and rapid oscillations may require dedicated instruments and an agreed power-quality or dynamic-study method. Reporting one value every second does not by itself prove that the instrument captures every sub-second event. Record the averaging window, filtering and bandwidth as well as the reporting interval. DOE’s May 2026 monitoring article specifically discusses why PMU filtering can obscure higher-frequency data-centre oscillations. [1]
3. Explain what creates the pattern instead of treating AI as a fixed multiplier
A facility can mix training jobs, interactive inference, storage, networking and conventional computing. The timing of their peaks depends on workload scheduling, hardware controls and how much activity is coordinated. Cooling and UPS behaviour can further change what reaches the utility meter. Request workload and equipment-state records so that an electrical event can be associated with a plausible operating cause.
Do not assume that every AI job can be paused or moved. Ask the operator which jobs can change, how long a change can last, how recovery works and what service or commercial limit applies. Until that authority is documented, record potential flexibility as unverified. This exercise teaches a measurement and calculation method using fictional data; it does not establish a validated AI load-forecasting model.
4. Separate a battery power requirement from an energy requirement
If facility demand exceeds an import ceiling, the required AC discharge at each instant is the positive part of demand minus that ceiling. The largest shortfall sets a preliminary power requirement. Integrating the shortfall over time gives the AC energy delivered during the event. These are separate checks: enough MWh cannot compensate for insufficient converter MW.
The installed DC battery must also account for conversion losses, ageing, protected SOC and repeated events. Recharging consumes import headroom and may compete with other loads. A simple peak calculation does not validate controller latency, ramp tracking, thermal limits, harmonics or outage reserve. Those issues need their own acceptance criteria and evidence.
5. Ask for data that allows another engineer to reproduce the result
Request timestamped demand at the utility point, IT feeders and relevant cooling feeders, with a shared time reference or a documented clock offset. State the study purpose before agreeing recording rates and duration with the electrical specialist. Include normal operation, credible workload transitions and recovery events; a quiet hour alone is not a representative acceptance dataset.
For each file, request meter identifiers, calibration or accuracy information, sampling and averaging settings, missing-data flags and the associated operating log. Agree how confidential workload records will be minimised and shared. Mark measured data, vendor estimates and fictional assumptions separately, and give every unresolved item an owner and a response date.
Workshop: form a decision from the given inputs
Attempt these tasks before opening the reference answers.
- Calculate the 15-minute average and explain whether it proves compliance with a 10 MW instantaneous import ceiling.
- Calculate the peak AC battery duty, the event energy and the average ramp rate in the separate transition record. Explain what each result can support.
- Calculate the AC recharge energy for one event, then write a conditional decision and a specific data request.
1. The average conceals an import breach
The interval energy is 8 × 14/60 + 14 × 1/60 = 2.1 MWh. Dividing by 0.25 hours gives 8.4 MW. The average lies below 10 MW, yet the final minute demands 14 MW and exceeds that instantaneous ceiling by 4 MW without battery support.
If a tariff bills interval-average demand, the billing result must follow that tariff. It cannot be used to reinterpret a separately agreed instantaneous connection constraint. A 15-minute meter value also cannot establish the shape of the hidden event.
The output from this step should be: A short decision note that records both interval average and instantaneous peak.
2. Power, energy and ramp answer different questions
Holding grid import at 10 MW during the plateau requires 14 − 10 = 4 MW AC discharge. The plateau energy is 4 × 1/60 = 0.0667 MWh, or about 66.7 kWh AC. These figures are an idealised duty at the chosen boundary, not the installed battery specification. A 2 MW converter cannot meet this duty even with a very large battery.
The separate record has an average ramp of (14 − 8)/2 = 3 MW/s. This is not proof of the maximum sub-second ramp, and it does not establish that a controller can track it. Because that record is separate, its transition energy has not been added to the ideal cycle. Real sizing requires integrating the actual measured trajectory.
The output from this step should be: A calculation sheet with units and explicit exclusions for controller and dynamic performance.
3. Recharge arithmetic supports only a conditional next step
Restoring the DC energy withdrawn requires 0.0667/(0.95 × 0.95) ≈ 0.0739 MWh of AC charging input. At a constant 2 MW this is about 2.22 minutes. The assumed 14-minute window offers 2 × 14/60 = 0.4667 MWh of AC input, so the ideal recharge opportunity is sufficient for one event.
The defensible decision is to investigate a battery solution, conditional on measured load trajectories, converter capability, protected reserve, repeated-event operation and an approved connection/control strategy. Request synchronised PCC and feeder data with instrument settings and workload logs, and ask the operator to approve any proposed workload adjustment. Do not promise a 4 MW / 0.0667 MWh purchased system from this exercise.
The output from this step should be: A data request identifying meter boundaries, instrument settings, operating events, permissions and accountable owners.
The key decision from this lesson
Choose the measurement boundary and time resolution before choosing equipment. An average can be correct while hiding the event that determines converter power or dynamic performance.
Sources and scope
Reviewed 25 October 2026. Primary sources provide background; numbered references in the text identify the relevant documents below. Cases, calculations and action suggestions are original teaching derivations. They do not replace site specifications, network permissions or contract review.
[1] U.S. DOE · Monitoring Oscillations from Large Data Centers · 28 May 2026 ↗