Reliability

The period and the conditions are half the number

Quote a reliability figure without a period and you have quoted nothing, because R falls monotonically with time. The same item can be 0.999 over a month and 0.988 over a year, and both are true statements about one λ (0.999 raised to the twelfth is 0.988).

Quote it without conditions and you have quoted something worse than nothing, because the conditions are what λ was measured or predicted under. A cabinet fan rated in a 25 °C laboratory is not the same fan inside an enclosure holding cells near 25 °C while rejecting heat to a 45 °C summer ambient, and the container thermal-management entry sets out how far those two environments diverge in practice.

The third quiet input is the definition of failure. Does a PCS that rides through a grid disturbance, latches out and needs a manual reset count as a failure? Does a module the BMS has isolated while the rack keeps running? Does a derate? On a modular battery plant a large share of events remove capability without removing the plant, so a reliability statistic assembled under a strict failure definition and one assembled under a loose one describe the same hardware with numbers that differ by an order of magnitude. Ask for the definition before you compare two vendors' figures.

Where the exponential form comes from, and where it stops holding

R(t) = exp(-λt) is not a general law of equipment. It is the solution you get when the hazard rate is constant — the assumption that the item is as likely to fail in its next hour of service as it was in its first. Only under that assumption does λ = 1/MTBF hold, and only under it can a single number stand in for a whole life.

Real equipment follows a bathtub hazard curve: a decreasing region where manufacturing and installation defects burn off, a roughly flat middle, and a rising wear-out region as bearings, contactors, capacitors and compressors reach the end of their design lives. Infant mortality and wear-out both violate the constant-hazard assumption, which is exactly why the exponential works well in the middle of life and badly at both ends.

One consequence of the exponential form is worth carrying around, because it corrects a common instinct. Survival to t = MTBF is exp(-1), about 37 percent — not 50 percent, and not a near-certainty. If the flat-hazard assumption applies at all, roughly two-thirds of a population will have failed at least once by the time the fleet has accumulated one MTBF each. That is the arithmetic that turns a large-sounding hours figure into an ordinary expectation of field events.

Reliability and availability pull apart in both directions

The two metrics answer different questions and can disagree sharply. A plant that fails often but recovers in an hour — a modular site with rack-level isolation, remote reset and spares on the shelf — can post excellent availability on a poor reliability record. A plant that fails almost never can post terrible availability if the one failure it does have takes six weeks to fix.

The availability entry makes exactly that point with a replacement transformer, and this site's transformer entries put the underlying lead times at roughly 12 to 18 months for MV units and 24 to 36 months or longer for HV station transformers. Nothing about component reliability rescues that.

The bridge between them is the steady-state identity the availability entry states: A ≈ MTBF / (MTBF + MTTR) for a repairable system. Reliability lives in the first term, restoration in the second, and the plant's number is the ratio. Which is why a spares strategy, remote diagnostics and a bounded repair unit usually move availability further than any specification you can write about a single component.

Why the contract stack buys availability and not reliability

Look through a BESS contract set and you will not find a guaranteed reliability figure. You will find performance guarantees in three families — retained capacity, round-trip efficiency, availability — with the availability guarantee at utility scale commonly landing at 95-98 percent measured annually, and mature LFP sites targeting 97 percent and above.

The reason is that availability is the thing the offtaker actually buys and the thing an EMS or SCADA historian can measure hour by hour, while reliability is a probability statement that cannot be settled from twelve months of one plant's data.

The one place a reliability-shaped test does appear is commissioning. Most contracts end with a reliability run: continuous or near-continuous operation over days to a few weeks against an availability threshold, with rules for which interruptions reset the clock. Passing it gates the commercial operation date. It is a separate number proven by a separate procedure from the operating-year guarantee, and carrying one into an argument about the other is a category error.

What reliability is actually for on a battery project

Reliability earns its place as a design input rather than a contract term. Components in series multiply: for independent items where any one failure takes the function down, R_sys = R1 × R2 × … × Rn, so system reliability is always lower than the worst element in the chain and never higher. Add a series item and you have subtracted reliability. That is the arithmetic behind every architecture decision on a battery plant — how many racks share a contactor, how many blocks sit behind one PCS, whether the main power transformer has anything behind it at all.

So the useful design question is not how to raise R but how to shrink the consequence of the events R predicts. Parallel redundancy, rack-level isolation, many small power blocks instead of one large one, and an auxiliary supply that does not take the site down with it. A plant built from thousands of repeated parts will generate events; the engineering choice is whether each event costs a slice or the site.