Power becomes energy
How much energy does a constant load request in half an hour?
kW and kWh
The power playbook / 24 practical lessons
Power Systems for Datacenter Engineers
A generator fails. The battery takes over. Can you predict when the lights go out?
Start with a 1 MW load and a 100 kWh battery, then change one assumption at a time. Read the calculations here, run each experiment in your browser, and compare the result with the reference Python engine.
Find your next question
The learning library
12 continuity lessons · 12 infrastructure lessons · 1 EMT study · 5 advanced studies.
Scenario presets are a separate collection.
No matching lessons or studies. Try a broader term, or clear the filters to see every path.
Continuity
12 course lessonsAccount for distribution losses
How much source power is needed to deliver 1,000 kW?
Efficiency
Calculate battery ride-through
Does half the stored energy give half the ride-through duration?
Finite storage
Bridge generator startup
What changes when startup takes 60 seconds instead of 30?
Sequence and delay
Add a failed generator
What changes if the generator cannot start?
Combined failures
Check a surviving path
Are two paths enough when either one must carry the whole load?
N+1 capacity reasoning
Expose shared control failures
Can a shared control fault defeat both power paths?
Failure domains
Find a single point of failure
What happens to two paths when their common main bus fails?
Common upstream assets
Track energy before the outage
Why does the 50 kWh case last longer when it can charge first?
Opening balances
Interrupt a 50 MW AI cluster
How long can 5 MWh of stored energy bridge a 50 MW load?
Aggregate AI power demand
Catch a sub-second service gap
Does 57 kWh bridge recovery at 500 seconds? Does 58 kWh?
Exact event timing
Separate PUE from continuity
How does assumed PUE change annual energy for the same 50 MW IT load?
Facility energy planning
EMT
1 separate studyInside a voltage sag
Explore a 200 ms DC-link transient. Change the circuit, replay the waveforms and check where the stored energy goes.
Power dynamics
5 advanced studiesLoad-step response
How quickly does PCC power follow a change from 0.5 to 0.6 pu?
Run an equilibrated 21-state converter chain. Compare the requested load, PCC response and DC-link voltage.
Modes and frequency response
What changes when the voltage-loop bandwidth changes?
Inspect eigenvalues, damping, dominant state participation and a sampled small-signal frequency response.
Synthetic load spectrum
Which part of a repeating load swing reaches the PCC?
Apply a bounded sinusoidal load and compare its time response and spectrum. The input is explicitly synthetic.
Phase model versus QSS
What does the reduced formulation leave out?
Compare the supplied 41-state abc and 21-state quasi-steady formulations, including their hand-set startup and a shared step at 0.04 s in a bounded 0.08 s window.
Interacting grid ports
Does a solved power flow imply stable dynamics?
Explore a modified nine-bus network with synchronous-machine, grid-forming, grid-following and data-center ports.
Infrastructure Engineering
12 course lessonsBuild an evidence-controlled load ledger
Keeping 6.2 MW cooling unchanged, can the revised design ratio accommodate the ledger?
Begin where the original course leaves off: not with a bigger battery, but with a more disciplined definition of demand. A facility total is meaningful only when every term belongs to the same boundary and operating condition. In this lesson, a 20 MW IT allocation becomes a traceable facility ledger, while a second example demonstrates why conflicting source values must remain visible.
Test capacity, paths and common dependencies
Does one spare still leave enough usable capacity?
The existing lessons show how a surviving path or shared controller can govern electrical service. Now extend that intuition into a state matrix. Installed equipment, connected equipment and independently substitutable equipment are different sets. A spare counts only when it can reach the required load in the state being assessed.
Declare transformer and fault-current boundaries
What changes when the upstream source is idealized as infinite?
A current value without its voltage, power base and network boundary is easy to misuse. This lesson builds a small equivalent circuit in layers, then shows how changing an assumption changes a fault-current screen. The purpose is to expose model limits before equipment selection, not to turn an attractive fault animation into a completed short-circuit study.
Audit generator banks and usable fuel
Does thirteen installed per side close the capacity screen?
The original startup lessons ask whether a generator becomes available in time. This lesson asks a different question: can the available bank carry the backed load at all? Follow a source claim through its real-power conversion, post-loss capacity and fuel ledger. A failed arithmetic claim should stay failed even when the combined installed nameplates look impressive.
Separate UPS power, battery energy and trip DC
How do the energy screens change at the same output power?
The battery lesson already distinguishes reserve from ride-through. Now separate three further boundaries: AC output energy, delivered DC energy and nominal end-of-life storage. Then examine the protection DC circuit, where enough ampere-hours can coexist with insufficient voltage at the device that must operate.
Trace protection zones and the full trip chain
What happens near the pickup boundary?
Protection is not a green status field at one current. It is a chain from the actual fault through measurement, logic, DC supply and interruption. Use a declared inverse-time curve to learn how a calculation behaves, then inspect what that curve cannot establish about CT saturation, electrical zones or selectivity.
Inspect grounding, distortion and control-power loops
What changes if the fundamental current falls while harmonics and reference demand stay fixed?
Low resistance, low distortion and a nominal battery voltage each answer only a narrowly defined question. This lesson turns those single numbers into explicit boundaries. Follow the ground current, compare harmonic denominators, and include the return conductor in the trip circuit before drawing any conclusion about service.
Coordinate liquid cooling and thermal ride-through
How much room heat remains at the new rack boundary?
Battery minutes do not create cooling minutes. Start with the manuscript’s mixed-cooling rack, then move to a separately labelled facility thermal model. The visual goal is to make both heat paths and their shared rejection boundary visible, so that the learner never mistakes a liquid-cooled label for complete removal of room heat.
Separate physical routes, optical budgets and fabric capacity
Does one uplink failure change the optical budget of the remaining channel?
Two fibres are not automatically two independent services, and a low optical loss is not automatically an application pass. Separate the physical administration route from the logical fabric, then calculate a narrow loss budget and post-uplink-loss capacity. The lesson adds communications reasoning without pretending that the existing electrical simulator already models packets.
Review spatial capacity and phased development
What average pressure does the new rack screen produce?
A rack count is not a facility plan. Equipment requires service space, replacement routes, structural capacity and coordinated boundaries through every construction phase. Use the manuscript’s historical operating account to practise reporting what was actually stated without turning different area descriptions into one misleading density metric.
Build an integrated commissioning evidence loop
Which teaching constraint fails when staged restoration takes 23 s?
A completed component test answers a component question. Integrated continuity requires an initial state, expected sequence, instruments, acceptance criteria, stop conditions and recovery boundary that span the service chain. Compare two clocks, then examine the measurement boundary behind an apparently precise efficiency number.
Audit the correction before releasing the design
Does correcting the transformer count close the whole design?
The final lesson asks for a justified release decision, not the largest possible green score. Bring capacity, connectivity, protection, thermal behaviour, controls and evidence together. Then audit a proposed correction from the manuscript’s separate campus example: a remedy is another engineering assertion and must be recalculated.
All five power-dynamics studies share one overview; each experiment opens its own configuration. Learning-content manifest and source hashes.
From system to explanation
See what each model can tell you
Illustrative relationships, not a replay of solver results. Open a lens to reveal a reviewed explanation. Only Run experiment opens the numerical model.
Continuity / system view
Trace a path. Account for the energy.
Utility, generator and finite battery supply feed distribution paths to the IT load. The path block aggregates the modeled topology; individual lessons expose parallel paths and shared failures. Availability and capacity determine what reaches the load.
Static diagram. No simulation is running.
Four lenses / select to explore
01 Physical
Utility, generator and finite battery supply feed distribution paths to the IT load. The path block aggregates the modeled topology; individual lessons expose parallel paths and shared failures. Availability and capacity determine what reaches the load.
02 Model
Power is measured in kW and energy in kWh. For the charging-disabled 1 MW example: 100 kWh × 0.90 × 0.95 ÷ 1,000 kW × 3,600 s/h = 307.8 s of ride-through. The outage starts at 300 s, so depletion occurs at 607.8 s elapsed. Lesson 12 uses a separate annual PUE ledger; it is not an outage simulation.
03 Verification
Compare the applied input and full result with the reference Python engine. The canonical evidence packet and lesson default/challenge records make software reproduction inspectable. Agreement between implementations does not calibrate a facility.
Read the guide, evidence and sources →04 Decision
Use the result to explain finite reserve and failed supply paths. It does not predict GPU jobs, grid adequacy, transient voltage or certified uptime. A real facility requires equipment data and qualified engineering review.
EMT / system view
Stored energy changes the voltage response.
An ideal DC source feeds a series resistor and inductor. A capacitor and constant-resistance load share the bus. This is a DC-equivalent teaching circuit.
Static diagram. No simulation is running.
Four lenses / select to explore
01 Physical
An ideal DC source feeds a series resistor and inductor. A capacitor and constant-resistance load share the bus. This is a DC-equivalent teaching circuit.
02 Model
L di/dt = Vsource − R i − Vbus; C dVbus/dt = i − Vbus/Rload. The source falls at 40 ms and recovers at 120 ms in a 200 ms study. Internal units are A, V, s, Ω, H and F. Integration steps and plotted samples are different.
03 Verification
Inspect the signed energy ledger, compare JavaScript with Python and halve the integration step. Analytic and step-refinement checks test the numerical implementation. They do not establish PSCAD equivalence or measured calibration.
Read the guide, evidence and sources →04 Decision
Explain how inductance, capacitance and resistance shape a voltage sag. The ideal source may absorb reverse current, and there is no diode or voltage clamp. Converter switching, protection and an actual facility are outside this circuit.
Power dynamics / system view
Follow states, modes and interacting ports.
The five studies explore converter chains, small-signal modes and a modified nine-bus network. This relationship map is not a single electrical wiring diagram.
Static diagram. No simulation is running.
Four lenses / select to explore
01 Physical
The five studies explore converter chains, small-signal modes and a modified nine-bus network. This relationship map is not a single electrical wiring diagram.
02 Model
For a linearized eigenvalue λ = σ + jω, frequency is |ω|/(2π) Hz and damping ratio is −σ/√(σ² + ω²). Per-unit values use the documented model bases, not an automatic conversion to MW. Converter and network formulations remain separate studies.
03 Verification
Read applied settings, solver completion, residuals and recorded tolerances. The owner-supplied Python studies were adapted; original MATLAB models and a measured GPU trace were absent. Historical v1.0.0 release assets predate these studies.
Read the guide, evidence and sources →04 Decision
A converged power flow does not imply dynamic stability. A local mode does not certify a whole operating envelope. The spectrum input is synthetic, the nine-bus case is modified, and none of these studies validates equipment or controls it.
Build your power-systems intuition
For datacenter engineers learning power continuity. Begin with power and energy, work through outages and shared failures, then compare aggregate AI demand and annual PUE planning. Basic arithmetic is enough to start.
Go inside a voltage sag
Ready for a different timescale? The separate EMT fundamentals study explores a 200 ms ideal DC-link RLC transient with waveforms, an energy ledger and step-refinement checks. Continue with five advanced power-dynamics studies of converter response and grid modes.
What the continuity lessons establish
These are deterministic synthetic electrical examples with explicit units, input scenarios and downloadable results. They teach reserve accounting and failure-path reasoning. They do not predict GPU jobs, grid adequacy, electrical transients or certified facility uptime.
Read about the author, model scope and reproducibility evidence →
Website source revision: 4878472c4d6c. The engine version does not identify every study model; see each guide and applied result.