Degradation curve
A degradation curve is the line a battery project ages along on paper: retained capacity on the vertical axis, time or cumulative cycling on the horizontal, sliding from 100% of some reference capacity toward the contractual End of Life floor.
It looks like data and is mostly assumption — every curve is conditional on a duty profile, a temperature history and an average state of charge, which is why no universal retention curve exists for a chemistry, a vendor, or even a single product.
Three different objects wear the same name — the warranty floor the supplier guarantees, the expected trajectory the financial model runs on, and the sparse measured points the capacity tests produce — and most degradation disputes begin with two parties each holding a different one. Reading a curve therefore starts with its axes and its footnotes, not its shape.
Reviewed August 2026 by Sergey Syrvachev
New to BESS? Start free with the 7-email fundamentals course — no cost, no account.
What it is (precise)
Start with the axes, because both are choices. The vertical axis is retained capacity as a percentage of a reference — and the reference must be named: beginning-of-life nameplate or the (often higher or lower) commissioning capacity test; DC energy at the racks or AC energy at a stated boundary; a single cell in a laboratory or the whole system with its weakest-rack and imbalance effects included.
The horizontal axis comes in three currencies — calendar years, equivalent full cycles, or MWh of throughput — and each embeds a duty assumption. A curve drawn in years silently assumes a cycle count per year; a curve drawn in EFC silently assumes a calendar-aging contribution that only holds at one dispatch rate and temperature. Two curves on different x-axes describe the same plant only under one specific operating pattern, and converting between them requires the duty profile that the plot itself does not show.
The curve is the superposed output of the two aging channels — calendar fade running on time, temperature and resting SOC, cycle fade running on throughput, depth of discharge and C-rate — evaluated at one operating point. The Degradation entry covers the mechanisms and the four distinct fades; what matters here is that the published line is almost always an energy-retention curve.
Power capability fades on its own trajectory as resistance grows, and a plant can track its energy curve faithfully while losing the ability to hold rated MW across a full discharge — a separate line that most curve exhibits never draw.
Warranty curve, expected curve, measured data
The warranty curve is a negotiated floor: a year-by-year retention table in the supply-agreement annex that the supplier is confident of clearing with margin, valid only inside an operating envelope of cycles or throughput per year, C-rate, temperature and SOC-dwell limits.
The expected curve — the P50 case — is a model output: cycle-test data extrapolated through the project's assumed site duty and climate, and it is what the augmentation schedule, the year-by-year revenue lines and the levelized cost of storage are actually built on. The measured data is neither of these: it is a handful of capacity-test points, typically annual, taken at the contract meter under the contract's test conditions, with interpolation doing the work between them.
Each object has one correct use. Model energy and revenue on the expected curve with sensitivity cases; enforce the floor when a test point breaches it; and treat the gap between the two as the supplier's engineering margin, not as free energy the buyer can bank. Lenders' independent engineers spend their time exactly here — benchmarking the vendor's expected curve against third-party degradation models and the cycle-test data behind it.
The uncomfortable zone is a fleet running below expected but above warranted: no remedy is triggered, yet every augmentation date in the financial model just moved earlier, and that cost lands on the owner. A model that carried only the warranty curve never sees it coming — and one that carried the warranty curve as its base case has quietly overstated fade and overbuilt the plant.
No universal curve exists, which is why this figure carries no numbers: the curve is a model evaluated at one duty point, and the same cell under arbitrage duty and under capacity duty traces two different lines. Typical LFP shape is about 2–4% fade in the first year and then roughly 1–2% a year under moderate duty, with a possible late-life knee that early data cannot rule out. What bends it: duty (EFC, depth of discharge, C-rate), sustained cell temperature — calendar fade roughly doubles per +10 °C — average and resting state of charge, and cumulative throughput. Normalize the end threshold before comparing anything: 10,000 cycles to 70% is a WEAKER claim than 8,000 to 80%.
- The axes
- y: retained capacity in % of a named reference (nameplate vs tested, DC vs AC); x: years, EFC, or MWh throughput — each embeds a duty assumption
- Three curves, one name
- Warranty floor (negotiated, envelope-conditioned), expected/P50 (model output the financial case runs on), measured capacity-test points
- Typical LFP shape
- ~2–4% fade in the first year, then ~1–2% per year under moderate duty — with a possible late-life knee that early data cannot rule out
- What bends it
- Duty (EFC, DOD, C-rate), sustained cell temperature (calendar fade roughly doubles per +10°C), average/resting SOC, cumulative throughput
- Why no universal curve exists
- The curve is a model evaluated at one duty point — the same cell under arbitrage duty and capacity duty traces two different lines
- End-threshold trap
- Cycle-life claims run to a stated floor: 10,000 cycles to 70% is a weaker claim than 8,000 to 80% — normalize before comparing
- Where it is verified
- Periodic capacity tests at the contract boundary, temperature-corrected — the only measured points; everything between is interpolation
- What it drives downstream
- Deliverable MWh per revenue year, augmentation trigger dates, LCOS, and the debt sizing built on all three
What bends the curve
Four levers set the trajectory. Duty — cycles per year, depth of discharge and C-rate — feeds cycle fade in proportion to how hard and how deep the plant works. Temperature feeds calendar fade, which roughly doubles for every 10°C of sustained cell temperature. Average and resting SOC compound it: time spent high and hot is the most expensive idling a battery can do.
Cumulative throughput ties the account together. For LFP under moderate duty the resulting shape is familiar — a steeper first year, often around 2–4%, settling toward a near-linear 1–2% per year — but the shape is an output of those inputs, not a property of the chemistry.
Two features deserve suspicion. The first is the knee: fade can accelerate late in life under high temperature, high average SOC or sustained lithium plating, so a straight-edge extrapolation of early near-linear data to year 20 is an optimistic act, not a neutral one. The second is the reference duty. Vendor cycle-life data is typically measured under continuous laboratory cycling at 25°C and moderate C-rate — no rests, no partial cycles, no seasons — while the site delivers all three; the Duty profile entry covers that gap in detail.
Run the same cell through an arbitrage duty and a rarely-dispatched capacity duty and you get two visibly different curves from identical hardware. That is the whole argument against a universal retention curve: the curve is a model evaluated at one duty point, and quoting it without the duty is quoting half a number.
How it shows up in specs, models and contracts
On a cell or DC-block datasheet the curve appears as cycle life: retention versus cycle count at a stated temperature, C-rate and depth of discharge, terminating at a stated end threshold. The threshold is part of the claim — 10,000 cycles to a 70% floor is a weaker promise than 8,000 to 80%, and claims quoted to different thresholds cannot be compared until normalized.
In the supply contract the curve becomes the capacity-warranty retention table plus the envelope that conditions it; in the financial model it becomes the deliverable MWh in every revenue year, the trigger dates in the augmentation plan, and a first-order driver of the debt the project can carry.
Verification is where the curve meets a meter. Periodic capacity tests — temperature-corrected, at the contractual boundary, under the contract's own protocol — produce the only points on the plot that are measurements rather than model.
They are sparse, so how the contract treats the space between them matters: whether shortfall is interpolated or only counted at test dates, how augmentation energy added mid-life enters the retention arithmetic, and which document's test governs when the warranty and the offtake disagree. Disputes about degradation are usually disputes about test method and reference conditions wearing a chemistry costume.
Questions that expose the assumptions
Any curve put in front of a buyer should survive six questions. Retention of what — nameplate or tested capacity, DC or AC, at which boundary? Against what x-axis, and what cycles-per-year figure converts it to calendar years? Under what duty — EFC per year, depth of discharge, C-rate, average SOC, cell temperature — and does that duty match the trading strategy the revenue model assumes?
Is this a guaranteed floor or an expected case? To what end threshold does the underlying cycle-life data run? And how does the curve treat augmentation energy once new racks join the count? A vendor who answers all six crisply has a model; one who answers none has a marketing chart.
The recurring failures are all comparisons that were never valid. Overlaying two vendors' curves drawn against different references, duties or end thresholds is not a comparison — normalize first or not at all. Feeding the warranty floor into the revenue model overstates fade and mis-sizes the overbuild; presenting the expected curve to a counterparty as if it were guaranteed does the reverse. And a curve that was true at signing stops being true the day dispatch changes: re-optimizing the plant into a harder duty re-derives the whole trajectory, whether or not anyone re-runs the model.
The degradation curve is a property of the battery — once the vendor publishes it, you can read year-10 capacity straight off the plot and put it in the model.
In reality: The published line is a conditional prediction, not a measurement of the product. It was generated under one reference duty — usually continuous laboratory cycling at 25°C — and it holds only for a plant operated at that duty, temperature and average SOC. Change the dispatch and the trajectory has to be re-derived. Worse, the plot in the warranty annex is not even the prediction: it is a floor the supplier expects to beat, so reading it as a forecast overstates fade, while reading the expected curve as a guarantee overstates certainty. Year-10 capacity is not on any published curve; it is the output of a model fed with the duty your plant will actually run — and the capacity test that eventually checks it.
Degradation curve, in context.
The Grid-Scale BESS course covers degradation curve — and the rest of the system — from the ground up, the way it actually gets deployed.