Performance normalization

Two legitimate directions, chosen in advance

There are two honest ways to compare a measurement taken at the day's conditions against a guarantee written at reference conditions. Correct the measurement to the reference conditions using published coefficients, or redraw the guaranteed figure at the day's conditions and compare there.

Either is defensible, and they are not equivalent in effort: the first produces one adjusted number, the second produces a moving target both parties have to accept. What is not defensible is picking between them once the raw numbers are in, which is the usual failure. The method, its data source, the reference conditions it targets and the acceptance band belong in the procedure both parties sign before the test window opens.

The third option is to test at reference conditions and skip the correction. On a real site that is mostly unavailable. Ambient is whatever the season supplies, the grid holds whatever voltage it holds, market windows decide when a full discharge is even possible, and a four-hour plant crosses its whole usable window in four hours, so there is no waiting for better weather mid-test. Normalization exists because reference conditions are a drafting convenience and the test day is not.

Temperature dominates, and it is cell temperature

Temperature moves the result through two mechanisms at once, and in the cold they push the same way. The energy available inside a fixed voltage window falls.

Internal resistance rises — electrolyte ionic conductivity drops by roughly an order of magnitude between +25 °C and −20 °C — so terminal voltage sags harder under the same current, the weakest cell in the string reaches its lower cutoff earlier, and the discharge terminates with charge still in the pack. Some of that energy is not lost at all; it reappears once the rack warms or rests. A cold test understates the plant twice, and neither effect is a defect.

Heat works the other way on the cells and against you at the meter. Warmer cells give up their energy more readily, but the thermal system holding them there draws more, and where station service is tapped on the plant side of the guarantee meter that draw sits inside the measured number. A POI-boundary capacity test therefore has a temperature optimum that is not the cell's optimum, and a correction applied on the cell side alone does not describe what the meter saw.

The trace to correct against is logged cell temperature from the BMS, not the site weather record — the same distinction warranty envelopes make, because a thermal-management fault can hold racks hot while the ambient reads normal. Record the distribution as well as the mean. A fleet whose hottest and coldest racks differ materially is not described by one number, and it is the extreme rack that ends the discharge.

The window, the rate, and the auxiliaries

The state-of-charge window in force at test time is a setting, not a property of the plant. A firmware change that moves a cutoff moves usable energy directly with no hardware touched, so the window the EMS actually enforced on the day belongs in the test record beside the temperature. Correcting a result measured on a narrower window to a guarantee written on a wider one is arithmetic anyone can do, and nobody should do silently.

Rate correction is the one most often skipped. Ohmic loss scales with the square of current, and the deeper sag brings the string to its floor sooner, so a discharge at rated power yields fewer megawatt-hours than the same window swept gently. When the guarantee is written at the contract duration and the test ran at something else — a market window closed, a block was out — the difference is real and needs a coefficient rather than an assurance that it is small.

Auxiliary allocation is not a physical correction at all. It is bookkeeping, and it decides borderline results. Whether HVAC, controls and thermal management sit inside the measured number depends on which side of the guarantee meter station service is fed from; if the guarantee is net at the POI and station service is fed from a separate grid-side supply, the auxiliary energy has to be deducted by calculation from a separate meter, and that calculation is an allocation both parties have to agree on.

Count the burn during the rest period between charge and discharge as well — it is real energy, and a procedure that only thinks about the discharge leg drops it.

The guarantee curve is dated

A retention guarantee is a schedule against a start date, so normalizing a result means comparing it against the right point on the curve. Which start date the curve counts from — shipment, energisation, or COD — decides which point applies, and the gap between those dates is where pre-COD degradation lives.

Calendar fade follows an approximately square-root-of-time trajectory, steepest at the beginning, so a three-month offset moves the year-one value far more than the year-twelve one. The rate is thermally activated, roughly doubling per 10 °C rise in cell temperature, so a hot laydown yard shifts the curve itself rather than only the reading taken against it.

Between scheduled tests the curve still has to be read somehow, and the contract should say how — interpolated between annual points, stepped, or ignored until the next anniversary. Tests rarely land on an anniversary. It matters most when the plant is close to its floor, because the same measured megawatt-hours can clear an interpolated curve and miss a stepped one.

Where normalization goes wrong

Double correction is the quiet failure. A vendor figure may already carry a correction — a usable-energy line quoted at a reference temperature, an efficiency number measured without the internal power supply — and subtracting the same effect a second time understates the plant as badly as skipping it overstates it. Before applying any coefficient, establish what the number it is being applied to already contains.

Treating the corrections as independent is the expensive failure. They interact. One converter family shows no DC-voltage derate at all below 40 °C and a clear knee at 50 °C; cell resistance is a surface over state of charge, temperature, current direction and age rather than a coefficient. Adding single-axis corrections computed at separate reference points produces a number the hardware will never deliver, in one direction or the other. Where an interaction matters and no coefficient covers it, the honest report says which part of the result rests on measurement and which on a model.

The last one is procedural, and it is the one that reaches arbitration. If the correction method was not named in advance, each party arrives with a formula that favours it and neither is wrong on its own terms. That is not a measurement disagreement, and no amount of retesting resolves it.