Defining renewable energy KPIs that survive real data

A KPI definition rarely fails loudly. It does not throw an error or leave a blank cell. It produces a number that is the right shape, in the right unit, on the right chart, and is wrong in a way nobody can see from the chart. That is what makes this class of failure expensive: the number is used, then used again, and the discovery happens months later when somebody tries to reconcile it against something else.

Five failure modes cover most of what goes wrong, and they recur across every industry and every asset class. None of them is exotic, and none is fixed by choosing a better formula. Each is fixed by deciding, in the definition itself, what the calculation should do when the world does not cooperate.

The common thread is that all five are about absence and change rather than about arithmetic. A KPI definition that only says how to compute the number is half a definition. The other half is what the number should do when an input is missing, when two inputs disagree about time, and when the past turns out to be different from what was recorded. Writing that half down is most of the work, and skipping it is why a KPI program that looked finished keeps producing surprises.

There are three tempting answers and all three are wrong. Dropping the point hides the outage and leaves a gap that later reads as no problem. Zero-filling fabricates data and quietly drags every average down. Publishing the number as though nothing happened is the worst of the three, because it is indistinguishable from a good number and therefore gets used. The correct answer is to publish the number and publish how complete its inputs were, as two separate facts.

Measuring completeness honestly is more subtle than counting samples. The denominator has to come from what each input actually promised to deliver, not from how often something happened to poll it: an input that reports only on change is bursty by design, and judging it against a poll rate would condemn a perfectly healthy signal. The numerator has to count occupied intervals rather than raw samples, or a publisher that sends a thousand readings in one minute and nothing for the rest of the hour scores as fully covered while the window is mostly dark.

When a calculation has several inputs, the grade is taken from the worst one rather than averaged across them. Averaging is the intuitive choice and it is the dangerous one: a fully sampled signal would mask a starved one, and the resulting number looks well supported while resting on a signal that was barely present. Taking the minimum means the grade describes the weakest leg the value stands on, which is the only reading of it that a person can safely act on.

Good

Inputs met the completeness threshold set on the definition. The value can be used as it is.

Uncertain

Inputs fell below that threshold. The value is still written, still charted, and still exported, carrying the grade so a reader can judge it. It is never silently dropped.

Unknown

At least one input states no delivery guarantee at all, so completeness carries no information. The grade says so rather than pretending to a number, and consumers branch on the grade rather than on the ratio.

One more rule belongs here because it is the one people most often get backwards. An asset with no observations at all in the window produces no row, rather than a zero. Zero is a measurement, and asserting one for an asset that reported nothing is fabrication. The same applies to a result that is undefined, such as a ratio whose denominator was zero: it is skipped, not filled. A gap in a chart is honest, and a fabricated zero is a claim about the plant that nobody made.

Because it removes the reader's ability to be appropriately suspicious. Given nothing, a person investigates. Given a number with no qualification, a person uses it, and the whole point of publishing a performance figure is that somebody acts on it. An ungraded number transfers the risk from the system, which knew its inputs were thin, to the reader, who cannot possibly know.

The effect compounds when one KPI is built from others, which is common and useful: a site score composed of several asset-level indicators, or a monthly figure rolled up from daily ones. Coverage alone launders quality here. A daily composite expects one value from each member per day, so a member that wrote a single thin, barely-supported value grades as fully covered. Nothing about the composite would reveal that every one of its inputs was uncertain.

The fix is that grades propagate rather than reset. A calculation that consumes other KPI values reads their grades as well as their numbers, and takes the worst one forward, so an uncertain input makes an uncertain result and an unknown input makes an unknown one. It is the same discipline as taking the minimum across raw inputs, applied one level up, and it is what stops a hierarchy of rollups from quietly manufacturing confidence at every level.

Because a name is not a definition, and everybody involved is acting in good faith. Availability is the classic example. One site excludes hours when the grid asked the asset to stop, on the grounds that the asset was capable and was told not to run. Another site includes them, on the grounds that the asset did not produce. Both are defensible, both are documented somewhere, and the two numbers are not comparable. Nothing on a dashboard says so.

• Does an externally requested reduction count against the asset?
• Does planned maintenance count as downtime, or is it excluded from the denominator?
• Does the day start at midnight local to the site, or midnight in a single reporting zone?
• Is a partially derated asset available, unavailable, or something in between?
• Which state is the asset assumed to be in before the first reading of the window?

None of these questions has a universally right answer, which is exactly why they have to be answered in the definition rather than left to whoever builds each site's version. The failure is not disagreement. The failure is undocumented disagreement, discovered when two sites are ranked against each other and one of them objects.

This is the practical argument for attaching a definition to an asset type instead of to a site. When there is one definition for the fleet, the conventions are decided once, by someone accountable, and every asset of that type is measured under them. Sites can still argue about whether the convention is right, which is a healthy argument to have in the open. What they cannot do is diverge silently and produce a comparison that looks valid and is not.

Time is where correct-looking definitions most often go quietly wrong, because every part of it seems obvious until a fleet spans more than one place. Daily is the first casualty. If daily windows are cut in one reporting zone, a site three hours away is being measured over a day that starts mid-morning and includes part of two of its own days. Daily aligned to each site's local midnight is almost always what the operator means, and it is a decision the definition has to carry rather than one the chart can fix later.

Accumulating counters are the second. Energy registers reset when equipment is serviced, roll over at their maximum, and start again at zero when a device is replaced. A plain difference across those events produces a spike or a large negative, and both are worse than a gap because they look like events. A definition that handles counters properly reads boundary values at the edges of each interval rather than summing differences, which also means a recomputation cannot double count, and it omits the value entirely across a reset rather than inventing a believable one.

Boundaries are the third and least obvious. A rate or a difference computed strictly inside a window has nothing to compare its first reading against, so the first interval of every window is either wrong or missing unless the calculation reaches back before the window starts for the last prior value. The same reasoning applies to state: a definition that measures time spent in a condition has to say what it assumes about the state before its first observation, and the honest assumption is usually that it does not know, so it does not start counting until it has actually seen the transition.

These are the three numbers a renewable portfolio is usually judged by, and their formulas take one line each. The formula is not what makes them hard. What follows is what each one asks of the data underneath, worked through the failure modes above, because that is where two sites reporting the same KPI stop agreeing with each other.

All three drift, for the same reason, and it is never carelessness. A capacity factor whose rated capacity was corrected after a repower, a performance ratio rebound to a replaced irradiance sensor, a round-trip efficiency moved from the cells to the connection point after a contract review: each change is legitimate, and each makes every value stored before it incomparable with every value after it unless the change was made as a new revision with a stated reach. Two sites reporting a round-trip efficiency without stating their boundary are not disagreeing about a battery. They are reporting different quantities under one name.

Every useful KPI definition is eventually edited. A convention is corrected, a threshold is retuned, a signal is rebound after a controller upgrade. The question is not whether that happens but what happens to the values already stored under the old definition, and an answer of nothing is the reason so many historical series cannot be interpreted a year later.

The mechanism that makes this tractable is versioning plus a stated reach. A change creates a new revision rather than overwriting the previous one, so a stored value can always be traced to the definition that produced it. The person approving the revision also decides how far back it should be reapplied, and that decision is recorded with the revision itself rather than inferred from a setting. Zero is a legitimate answer, and it means apply this going forward and rewrite nothing. What is not legitimate is a change whose historical reach nobody chose.

Late data needs the same treatment for a different reason. A site that lost its link and reconnects delivers hours of buffered readings, and every window they fall into was computed over inputs that have since changed. A definition is only trustworthy if already computed windows are revisited on a cadence and replaced when their inputs have moved, and if that replacement is idempotent, so recomputing a window that has not changed produces the same value rather than a duplicate.

Two smaller properties matter more than they sound. The first is that preview and production run the identical calculation, so what a reviewer approved is exactly what will be computed, rather than an approximation of it. The second is that a definition the engine cannot run is surfaced as a refusal rather than skipped quietly. A definition that stopped producing values and told nobody is the most dangerous state in the whole system, because every downstream chart keeps rendering its last known point and looks alive.