Is high availability meaningless, then?

No — low availability is always bad news, and the metric remains useful hygiene. The asymmetry is the trap: high availability is fully compatible with severe energy loss, so it can clear an asset that energy accounting would flag.

What metric should anchor an O&M performance guarantee?

Weather-corrected Performance Ratio or an expected-versus-actual energy attainment ratio, with the measurement methodology — irradiance source and sensor class, metering point, exclusions, correction method — specified in the contract per IEC 61724-1.

How quickly can an owner see their real gap?

With interval telemetry and a credible weather reference, a trailing-year energy reconciliation takes weeks, not quarters. Vector produces it as a standing view. The first one is usually the most expensive report an owner ever reads — and the most valuable.

Doesn't string-level detection require new hardware?

Usually not. Most modern string inverters already report per-MPPT or per-string currents; the gap is analytical, not instrumental. Where telemetry truly stops at the inverter, peer-deviation analysis still recovers most of the detection value.

How do we transition an existing availability-based contract?

Add the loss-attribution reconciliation as a reporting obligation first — it requires no commercial renegotiation and builds the shared evidence base. Move the guarantee itself to energy-based terms at the next renewal, once both parties trust the measurement.

The Availability Illusion: Why 99% Uptime Can Still Mean 8% Lost Revenue

Virat Damania — Co-Founder & Managing Director, Ellume Technology

Why 99% contractual availability can coexist with 8% revenue loss — and why the KPI at the center of your O&M contract is structurally blind to the way solar actually fails. Availability measures whether equipment is technically operational, not whether it is producing what it should. A plant can report 99% availability while dead strings, derated inverters, soiling, and tracker faults quietly remove 5–8% of producible energy — because availability is a time-based metric and energy loss is not. Owners should contract and manage to energy-based performance with full loss attribution, not availability alone.

Key Takeaways

→Availability was inherited from thermal generation, where units are binary; solar fails gradually, partially, and beneath the thresholds time-based metrics can see.
→The three structural blind spots: granularity (energy is lost at string level, availability is measured at inverter level), thresholds (derating and partial faults never trip 'offline'), and incentives (operators rationally manage to the metric the contract pays on).
→In a representative month, the only loss availability caught was the smallest one on the ledger — two brief inverter trips — while dead strings, a chronic underperformer, and soiling ran unmeasured.
→The fix is subordination, not abolition: keep availability as hygiene, govern with weather-adjusted energy attainment and a monthly loss-attribution reconciliation both parties can see.
→Attribution also protects good operators — it separates the losses that were theirs to prevent from weather, curtailment, and design constraints that never were.

Here is a plant I will not name. Sixty inverters, all online. Contractual availability for the month: above 99%. The O&M report was green, the invoice was paid, and everyone moved on. The same month, our physics model reconciled the plant's metered output against measured irradiance and found nine inoperative strings, one inverter producing at roughly half its fleet-peer average — and it had been doing so for months — and a combiner-level fault cluster concentrated in a single block. Together: roughly six percent of the month's producible energy, or in this plant's case about 16,000 kWh and $1,777, gone while every piece of equipment involved counted as 'available'. This is not an anecdote about a bad operator. It is a structural property of the metric. Availability was inherited from thermal generation, where a unit is either dispatching or it is not, and the binary question is the right question. Solar fails differently: gradually, partially, at the DC level, and beneath every threshold a time-based metric can see. We have built an industry's contracts, reports, and incentives on a KPI designed for a different technology's failure modes — and then we act surprised when portfolios miss their production estimates while every monthly report glows green.

Why is availability structurally blind?

Three reasons, and they compound. Granularity: availability is typically measured at the inverter or plant level, but solar loses energy at the string and module level. A central inverter with a quarter of its DC field dark is fully 'available' — it is online, converting, and reporting. On a representative 3 MW plant in our fleet with 240 strings, the string-level census read 187 healthy, 28 in warning, 16 critical, and 9 inoperative. The availability figure for that plant contained no trace of that distribution. Thresholds: derating, clipping-shift from degradation, thermal management faults, and tracker drift all reduce output without ever tripping the offline state that availability counts. An inverter running at 52% of its peer output for a quarter registers exactly the same availability as its healthy neighbors. And incentives: when the contractual KPI is availability, the rational operator manages to availability — hard outages get closed fast, because they are measured, while partial losses that no metric captures get no truck roll at all. The contract is working exactly as written. That is the problem.

What does the gap look like in numbers?

Read the last row twice. The only line availability caught was the smallest loss on the list. Everything material happened inside the 'available' state — which is exactly where a time-based metric, by construction, cannot look.

A representative month: what availability reports versus what energy accounting finds
Signal	Availability view	Energy attribution view
9 dead strings across 3 inverters	Invisible — inverters online	≈2.8% of monthly energy
1 inverter at ~52% of peer output	Available	≈1.4% of monthly energy
Soiling past cleaning trigger	Not an availability event	≈1.5% of monthly energy
Two 45-minute inverter trips	0.2% unavailability	≈0.2% of monthly energy
Reported month	99.2% available	≈5.9% energy below producible

How Ellume Vector makes the invisible losses visible

Vector was architected around the premise of this essay: that the governing question is not 'is it online' but 'is it producing what physics says it should.' The platform computes expected generation from measured irradiance and temperature at every interval, at plant, inverter, and string granularity, and attributes every deviation to a named, priced cause.

•The executive dashboard leads with an Energy Lost card carrying a dollar value — on the plant above, 16,156 kWh and $1,777 for the period — because a loss with a price attached gets managed, and a loss without one gets archived.
•The string health heatmap renders all 240 strings — every inverter on the rows, every string position on the columns, inoperative strings in black — so the nine dead strings that availability could not see are the first thing a morning glance lands on.
•Peer benchmarking flags the chronic underperformer: the affected inverter reading 48.5% below its fleet average, with the deviation, the duration, and the accumulated energy cost attached — not an alarm, a case.
•Every detected event expands into a four-part record — summary, raw telemetry readings at the moment of detection, the physics rule that fired with measured-versus-expected values, and a recommended action with effort, time-to-fix, and confidence — so the O&M conversation happens over shared evidence rather than competing recollections.
•The composite Plant Health score decomposes across seven dimensions (inverter health, DC string balance, grid stability, thermal margin, soiling, degradation, telemetry quality), which is how a plant can be honestly rated 'needs attention' at 80/100 while its availability reads 99% — the DC string balance dimension was sitting at 20.

From the Ellume fleet: the three critical string outages on that plant — INV-044, INV-051, INV-057 — all traced to the same combiner row in one block. Vector clustered them, with a pending soiling job, into a single dispatch: four problems, one truck roll. The same month, the platform's false-positive suppression avoided five unnecessary dispatches, saving $3,500. Availability-based management would have generated neither the detection nor the efficiency: the strings weren't 'down', and the nuisance alarms would each have rolled a truck.

What should replace availability?

Not replace — subordinate. Availability remains a useful hygiene metric, and low availability is always bad news. But the governing framework for an owner should be energy-based: expected generation computed from measured weather, actual generation against it, and the difference attributed to named causes with dollar values. From there, three contractual instruments follow naturally. Performance guarantees written on weather-corrected energy attainment or PR — with the measurement methodology specified per IEC 61724-1 — rather than on time. Reporting obligations that include a monthly loss-attribution reconciliation, so 'green' means accounted-for rather than merely online. And a shared recoverable-loss ledger that both owner and operator see, because the fastest way to align an O&M relationship is to make the unrecovered energy visible in dollars, every month, in the same report the availability number lives in. Owners sometimes worry this adversarializes the relationship. Our experience across the fleet is the opposite: the best operators embrace attribution, because it also exonerates them — it separates the losses that were theirs to prevent from the weather, the grid curtailment, and the design constraints that never were. When the ledger showed that plant's environmental losses were unavoidable and its fault losses were concentrated in one combiner row, the operator got a precise work order instead of a vague complaint. Blindness serves no one. It only feels comfortable.

Frequently Asked Questions

Is high availability meaningless, then?: No — low availability is always bad news, and the metric remains useful hygiene. The asymmetry is the trap: high availability is fully compatible with severe energy loss, so it can clear an asset that energy accounting would flag.
What metric should anchor an O&M performance guarantee?: Weather-corrected Performance Ratio or an expected-versus-actual energy attainment ratio, with the measurement methodology — irradiance source and sensor class, metering point, exclusions, correction method — specified in the contract per IEC 61724-1.
How quickly can an owner see their real gap?: With interval telemetry and a credible weather reference, a trailing-year energy reconciliation takes weeks, not quarters. Vector produces it as a standing view. The first one is usually the most expensive report an owner ever reads — and the most valuable.
Doesn't string-level detection require new hardware?: Usually not. Most modern string inverters already report per-MPPT or per-string currents; the gap is analytical, not instrumental. Where telemetry truly stops at the inverter, peer-deviation analysis still recovers most of the detection value.
How do we transition an existing availability-based contract?: Add the loss-attribution reconciliation as a reporting obligation first — it requires no commercial renegotiation and builds the shared evidence base. Move the guarantee itself to energy-based terms at the next renewal, once both parties trust the measurement.