Asset managers overseeing grid-scale BESS investments need to answer three questions over the project lifecycle: How much capacity does the system have today? How much will it have in three years? And when should we plan for augmentation or replacement? The answers drive revenue projections, PPA renegotiations, and repowering capital planning. Yet most RUL forecasting approaches in the field rely on manufacturer degradation curves that don't account for actual operating conditions — and that gap creates systematic planning errors that compound over time.
This article examines what bankable remaining useful life forecasting requires from the degradation model underneath, what the P10/P50/P90 framing means in practice, and where commonly used approaches fall short.
What "Remaining Useful Life" Means for a BESS Asset
RUL in the BESS context is typically defined relative to a minimum SoH threshold — most often 80% or 70% of original capacity — below which the asset is considered to have reached end of commercial life for the purposes of the original project contract. An asset with a 10-year warranty and a 70% EoL threshold has reached end of life when its measured capacity falls below 70% of its commissioning nameplate.
That 70% threshold is not a physical cliff. Batteries can and do continue operating below it — with reduced revenue per cycle and, in frequency regulation applications, reduced ability to meet dispatch ramp requirements. The threshold is a contractual and economic boundary, not a failure mode. Understanding this distinction matters for asset management because it affects how you model the value of the asset in the years approaching the threshold versus the years beyond it.
RUL forecasting answers: given the system's current SoH and its observed degradation trajectory, at what date will it cross the 70% (or 80%) threshold with a given probability? The "with a given probability" part is what makes probabilistic P10/P50/P90 forecasting useful rather than a single deterministic number.
P10/P50/P90: What the Probability Bands Mean
P50 is the median forecast — 50% of modeled scenarios reach end of life before this date, 50% after. P10 is the optimistic case — only 10% of modeled scenarios reach EoL before this date (meaning 90% degrade faster). P90 is the conservative case — 90% of modeled scenarios reach EoL before this date.
For asset management planning:
- P50 is used for central revenue projections and expected augmentation timing in base case financial models
- P90 is used for worst-case cash flow stress testing and for sizing augmentation reserves in project finance structures
- P10 is used for best-case projections in sale documentation or PPA extension negotiations where the seller wants to demonstrate upside
The width of the band between P10 and P90 reflects forecast uncertainty. A model with high uncertainty shows a P10–P90 span of several years; a model with good observational data shows a tighter band. For a newly commissioned system with only 6 months of operating data, a 3–4 year span between P10 and P90 at a 5-year horizon is typical. For a system with 3+ years of clean telemetry and regular capacity tests, that band should compress to under 18 months at a 3-year horizon.
Why Manufacturer Degradation Curves Aren't Enough
Battery manufacturers provide degradation curves as part of the product specification — typically a graph showing capacity remaining as a function of cycle count and calendar time under standardized test conditions. These curves are generated under controlled lab conditions: fixed temperature (usually 25°C), fixed C-rate (often C/5 for calendar tests, C/1 for cycle tests), and a defined SoC range (usually 0–100% or 20–80%).
Grid-scale BESS operate under conditions that deviate from these test conditions in ways that significantly affect degradation rate:
- Temperature variability. A BESS in Phoenix, Arizona operates at ambient temperatures 15–20°C higher than a BESS in Boston — and calendar aging rate roughly doubles for every 10°C increase in average temperature. The manufacturer's calendar aging curve, calibrated at 25°C, understates aging for warm-climate deployments by a factor of 2–4×.
- Partial cycling. Frequency regulation applications cycle the battery dozens of times per day at shallow depths — 5–15% SoC range per cycle. Manufacturer cycle life curves are calibrated at full-depth cycles. Shallow cycling is generally less damaging per cycle, but the high cycle count and the specific SoC window matter for lithium plating risk in some chemistries.
- Depth of discharge variance. Some dispatch profiles require occasional deep discharges below the typical operating range. These accelerate degradation disproportionately relative to the average depth of discharge.
- Thermal gradients within the rack. Cells at different positions in a rack age at different rates due to temperature gradients — even if the average module temperature is within spec, cells in hotter positions accumulate calendar aging faster than cells in cooler positions.
A physics-informed RUL model accounts for these factors by incorporating actual operating telemetry — measured temperatures, measured cycling profiles, measured coulomb throughput at the cell level — into the degradation projection rather than relying on standardized assumptions. The result is a forecast that is calibrated to the specific asset's history, not to a generic chemistry profile.
The Data Requirements for Reliable RUL Forecasting
The quality of a RUL forecast is bounded by the quality and completeness of the historical operating data. Minimum requirements for a forecast that narrows the P10–P90 band to operationally useful width:
| Data Type | Minimum Resolution | Minimum History | Why It Matters |
|---|---|---|---|
| Cell voltage time series | Per cell, ≤60 sec | 12 months | Capacity estimation, cycle counting, OCV-SoH mapping |
| Cell temperature time series | Per module, ≤60 sec | 12 months | Calendar aging acceleration factor calibration |
| String current | Per string, ≤60 sec | 12 months | Cycle counting, coulomb throughput, C-rate profile |
| Formal capacity tests | Per rack, C/5 protocol | 2+ tests, 6+ months apart | Ground truth SoH anchors for model calibration |
With less than 12 months of history, the model cannot separate calendar aging from cycle aging with statistical confidence, and the P10–P90 band remains too wide for meaningful capital planning. Twelve months of clean data is the minimum; 24–36 months produces substantially better forecast accuracy, especially for distinguishing linear from accelerating degradation trajectories.
Interpreting RUL Forecasts for Asset Decisions
RUL forecasts produce a predicted EoL date range. Converting that to an asset management decision requires combining the forecast with the project's financial and contractual structure:
Augmentation planning. If P50 EoL is in month 84 of a 10-year PPA, the project needs augmentation planned for year 7. Lead time for battery augmentation procurement — cell ordering, BMS expansion, grid interconnection approval amendments — typically runs 12–24 months. P90 EoL (the conservative case) should be the trigger for starting augmentation procurement, not P50. Using P50 as the trigger risks a 50% probability of falling short before augmentation arrives.
PPA renegotiation. When a PPA reaches its original term, the asset manager needs to demonstrate the system's residual capacity to support extension or renegotiation. P50 and P10 forecasts, backed by auditable cell-level telemetry and IEC 62619-format capacity test records, form the technical evidence package for that negotiation. A forecast unsupported by cell-level data carries much less weight in a counterparty negotiation than one backed by 3 years of clean telemetry and multiple capacity test data points.
Repowering vs. replacement. When P90 EoL is approaching and augmentation economics are unfavorable (cell costs relative to the original project's financing structure, remaining PPA term, etc.), the asset decision may be full repowering with a new BESS technology. RUL forecasts don't answer this question alone — they're one input into a broader repowering IRR analysis — but the accuracy of the capacity fade projection directly affects the IRR calculation's reliability.
In our work building degradation models for fielded systems, we've found that the most common planning error is using P50 for procurement trigger decisions rather than P90. The cost of getting this wrong — emergency augmentation procurement, missed PPA dispatch obligations, renegotiation from a weak position — is almost always higher than the cost of procuring augmentation capacity 6–12 months earlier than strictly necessary.
What Good RUL Forecasting Looks Like in Practice
A RUL forecast that's operationally useful has three properties: it is grounded in observed operating data rather than manufacturer assumptions, it provides explicit uncertainty quantification (the P10/P50/P90 bands), and it is updated continuously as new operating data arrives.
Static forecasts — a single projection produced at year 2 and consulted at year 5 — accumulate error as the system's actual operating profile diverges from whatever assumptions drove the original projection. Dynamic forecasts that incorporate new capacity test results and update their calibration as the telemetry history grows are more reliable precisely because they're continuously correcting against ground truth.
For asset managers responsible for storage projects with 10–20 year contractual lifespans, the difference between a static manufacturer-curve forecast and a dynamically calibrated physics-informed forecast is measured in the confidence interval around the augmentation timing decision. Get the timing wrong by 18 months and the cost difference — between planned procurement and emergency procurement, or between a strong PPA renegotiation position and a weak one — typically exceeds the total cost of implementing better monitoring infrastructure by a wide margin.
