METHODS / MODEL V0.1
Make the assumptions
as visible as the result.
This first study engine asks a bounded question: can advance workload signals improve a stabiliser’s energy readiness relative to a reactive policy, with identical plant constraints?
What stays the same
Both runs use the same demand trace, engine count and rating, aggregate ramp rate, stabiliser power and usable energy, conversion efficiency, energy-recovery setting and initial state. Both begin with generation equal to initial demand, capped at plant capacity. Neither policy intentionally reduces useful workload.
What changes
The reactive policy follows present demand and restores stored energy toward its initial value. The signal-aware policy uses the earliest announced future event within its look-ahead window to adjust that target. A positive expected load step prompts charging; a negative step prompts discharge to create absorption room. The policy does not inspect future demand rows.
Declared signals are assumed inputs, not predictions made by Mithra. Their notice times, expected changes and imposed delay are visible. Missing or too-late signals provide no advance action. In the no-signal case the two policies are identical.
Equations and time stepping
capacity = engines × engine_MW
recovery_MW = clamp((target_MWh − stored_MWh) × 3600 / recovery_seconds)
generation_command = clamp(demand_MW + recovery_MW, 0, capacity)
generation changes no faster than the shared aggregate MW/s limit.
estimated step-mismatch energy = expected_delta_MW² / (2 × ramp_MW_per_s × 3600)
The reserve target accounts for one-way conversion efficiency,
is bounded to 5–95% of usable energy, and uses a 5% energy margin.
discharge_MW ≤ min(deficit, buffer_MW, stored_MWh × efficiency / interval_hours)
charge_MW ≤ min(surplus, buffer_MW, remaining_MWh / efficiency / interval_hours)
generated + initial_storage = served + surplus + conversion_loss + final_storageIntegration intervals are at most 0.25 seconds and split at workload row boundaries. Generation changes linearly within an interval, with its average used for energy accounting. The demand trace uses zero-order hold. The final timestamp ends the run; its power value has no duration. Storage is treated as a bounded energy bucket. The engine continues through unmet demand and unabsorbed surplus; it does not simulate protective trips.
How to interpret the metrics
Unserved demand is demand the model could not supply. Unabsorbed surplus is generation the load and storage could not accept; no physical dump-load device is assumed. Buffer throughput is energy charged plus discharged at the bus. Conversion losses follow the selected one-way efficiency in each direction.
Generated energy alone is not a savings metric. Check unmet demand, losses and final stored energy together. This model contains no fuel curve, minimum stable engine load, engine start/stop scheduling, maintenance model or compute completion model. No fuel, financial or avoided-emissions benefit is calculated.
What this model cannot decide
It does not model frequency, voltage, governor dynamics, alternator inertia, protection coordination, start/stop sequences, thermal constraints, ageing or network topology. It cannot qualify equipment or determine safe protection settings. It is a starting point for a scoped engineering study and a testable software architecture.
Bring a workload
time_s,power_mw,event_delta_mw,notice_s
0,85,,
90,85,,
180,120,35,90
360,85,-35,270
600,85,,Required headers are time_s,power_mw. The optional signal columns must appear as a pair. Times start at zero and increase strictly; power is nonnegative in MW. An event signal refers to the time on its row, while its notice can be earlier. CSV import supports quoted fields, UTF-8 BOM and Windows line endings. Limits: 2 MB, 20,000 data rows, 2 hours and 100 signals. Record measurement origin, sampling and preprocessing in the provenance field. “User supplied” is not a verification of measurement quality.
Reproduce and review
A replay bundle contains normalized input points, declared signals, all plant settings, model/schema versions, a canonical SHA-256 input checksum and expected metrics. Importing it recalculates both runs and checks the checksum and metrics. The checksum detects changes; it does not authenticate the source. A source-matched Node command can run the same engine without the interface.
The HTML study report is self-contained, print-ready and includes assumptions, metrics, limitations and the checksum. The paired CSV exports every integration interval. Preserve the report, replay bundle, time series and matching model source together.
Data handling
Workload and replay files are processed in this browser. The application does not upload them, send their contents to a model, or store them in browser storage. Refreshing the page clears the current study; download a replay bundle to keep it. Website hosting receives ordinary page requests, and fonts load from Google Fonts. No analytics or customer-data backend is included.
From research to a plant-specific study
Agree the existing controls and the decision to improve. Obtain equipment limits, calibrated models or time-aligned measurements. Separate calibration and holdout cases, define acceptance criteria and include missing signals, equipment constraints and uncertainty. Review the results with qualified plant engineers before considering a separately scoped read-only pilot.
Background reading
- Power Stabilization for AI Training Datacenters — published research on workload variability and mitigation approaches.
- Bergen Engines: behind-the-meter power for AI data centres — the existing hardware context.
- Mithra controller foundation — the separate open-source interruptible-compute controller.
These references provide context. They are not validation of the parameters or results in this workspace.
Return to the study workspace ↗