Remaining useful life, from a measured trend to a real production cost

Remaining useful life is an estimate of operating time before an asset reaches a defined end-of-life criterion. Three inputs sit behind it, and all three carry uncertainty.

A measured degradation trend, normalised to reference conditions. A projection of how the plant will be operated between now and then. And the criterion itself, which has to be stated because different criteria give materially different answers from the same measurements.

None of the three is observable. The trend is inferred from readings taken under varying conditions, the projection is a forecast, and the criterion is a commercial or contractual choice rather than a physical fact. A figure that arrives without them attached cannot be checked, which is why a technical advisor will discount it.

For an electrolyser stack there are at least three defensible definitions of end of life, and they do not coincide.

A voltage threshold, usually expressed as a percentage rise in cell voltage at a reference current density. This is the definition most supply contracts use, because it is measurable and attributable.

An efficiency threshold, where specific energy consumption rises to a point that the plant's economics no longer work at the prevailing power price. This one moves with the power price, so the same physical stack has a different end of life in a different tariff environment.

An economic replacement point, where the marginal cost of continuing exceeds the annualised cost of replacing. This is the definition an owner actually cares about and the hardest to write down, because it depends on the replacement price, the outage cost and the remaining term of the offtake.

State which one the number refers to. A stack with four years to a voltage threshold may have two to an economic replacement point, and quoting the first while the model assumes the second is how a major maintenance reserve ends up wrong.

Cell voltage depends on current density, temperature and pressure, and it does not depend on them linearly. A degradation rate computed from raw readings taken across a varying operating envelope is dominated by the envelope rather than by the degradation.

The correction needs a reference. A polarisation curve captured at commissioning under controlled conditions, and repeated periodically, is what makes the normalisation defensible rather than assumed. Without it, the correction is being made against a supplier curve for a nominally similar stack, and the uncertainty in that substitution can be the same size as the degradation being measured.

The same reasoning applies to attribution. A plant-level efficiency figure includes conversion and auxiliary terms that move with load, so a rise in kWh per kg is not evidence of stack degradation until those are separated. That separation is set out in why a plant reads more kWh per kg than its datasheet, and it has to happen before a trend is fitted.

A degradation trend measured over eighteen months describes eighteen months of a particular operating pattern. Projecting it forward assumes that pattern continues, and on a renewable-coupled plant it will not: the resource varies year to year, the tariff structure changes, and the operating strategy is usually still being tuned.

The honest treatment is to project under stated scenarios rather than under one. A base case at the observed duty, a case at the contracted or budgeted duty, and a case at the duty the plant would run under a materially different dispatch pattern. Where the three converge, the estimate is robust. Where they diverge, that divergence is the actual finding and is worth reporting instead of averaging away.

Cycling deserves separate treatment, because its effect is not proportional to running hours. A plant that stops daily accumulates a different kind of wear from one that runs continuously at the same average load, and most supply contracts define a cycle in a way that matters for the count.

All three inputs carry uncertainty, and they compound. An RUL quoted as a single figure presents more confidence than the method supports, and an experienced technical advisor treats an unbounded number as a marketing output rather than an engineering one.

A defensible report gives a central estimate with a range, the criterion it was computed against, the duty scenario it assumes, the measurement basis and its uncertainty, and the date of the last normalised reference measurement. That is five lines, and it converts a claim into something someone else can audit.

It also ages. An RUL is a statement about a moment, and it should carry the date it was computed and the interval before it is recomputed. A figure carried forward unchanged through three reporting periods has stopped being an estimate.

Stack replacement is usually the largest recurring expenditure an electrolyser plant faces, and its timing is sized into the major maintenance reserve at financial close, from OEM curves and comparable projects rather than from the plant itself.

Once the plant runs, the measured trend is what confirms or corrects that assumption. A stack degrading more slowly than assumed is an argument for revisiting a reserve that is over-funded and constraining distributions. One degrading faster is a problem considerably cheaper to raise early than to discover at replacement.

Both directions cost money, which is the point worth making to a lender: this is not an exercise in arguing for a smaller reserve. It is an exercise in replacing an assumption made before the asset existed with a measurement taken from the asset itself. How that reporting is structured after commercial operation is covered in what lenders need after commercial operation.

Traceability. Every figure in the report should be followable back to measurements, and the measurements should be system-generated and timestamped rather than transcribed.

A stated method. Which normalisation was applied, which points were excluded and why, and what curve was fitted. Exclusions are legitimate and routine, and undeclared exclusions are what destroy credibility when someone else runs the numbers.

Cell-level data where it exists. A single stack voltage hides the distribution, and a population with a few outliers has a different remaining life from one degrading uniformly, even at the same mean. The two also have different remedies, since a small number of failed cells may be addressable without replacing the assembly.

Independence of provenance. Where the party producing the estimate is the party whose performance it describes, the estimate is still useful but should be labelled as such. Data that is system-generated and traceable does not remove that, but it does mean the numbers can be checked.

Yunify holds the normalised record and attributes movement to a mechanism rather than to a suspicion, which is what allows an RUL figure in a quarterly pack to be traced back to something physical. The estimate remains an estimate, and the criterion remains a commercial choice.