How to Assess BESS Degradation Modeling: What a Semi-Empirical Model Contains (Arrhenius Calendar Term, Square-Root and Linear Cycle Terms, SOC Dependence, Separate Mechanisms), How Well It Fits (1.4% RMS Capacity Error Across 0–55 °C; Below 1% End-of-Test Error on an LFP Storage Profile), What the Test Data Behind It Can and Cannot Say (BLAST-Lite's Models: 1–3 Years of Accelerated Aging Extrapolated to 10+; Sandia: LFP Cells Cycled 2,500–9,000 EFC, Most Not Yet at 80% Capacity When the Study Ended, Versus 200–2,500 EFC for NMC), Pack Versus Cell (BLAST-Lite's 20–30% Rule of Thumb; Cluster Energy Loss From Cell Inconsistency of up to 5–15% in Mid-to-Late Operation, per a 2026 Consultation Draft's Explanatory Note), and the Test and Standard Hooks (Reference Performance Test, IEC 62933-2-1, GB/T 36276 / 34131 / 36558 / 36549)

Short answer. NREL's 2017 paper says industry mainly uses semi-empirical lifetime models that extrapolate accelerated aging test data; the documentation of NREL/NLR's open-source BLAST-Lite models says those models are trained on 1–3 years of data and extrapolated to predict real-world lifetimes of 10+ years. The questions that decide whether such a model can be trusted are the same whoever built it. Structure: does it separate calendar aging (temperature, SOC) from cycle aging (throughput, rate, temperature) and track more than one mechanism? NREL's grid-storage model does this with 8 state variables for separate fade mechanisms; the TUM/NREL LFP model of Schimpe et al. models one calendar and several cycle effects, with an Arrhenius temperature law for calendar aging referenced at 25 °C and 50% SOC, a Tafel-type SOC dependence built on the anode potential, square-root high- and low-temperature cycle terms plus a linear low-temperature/high-SOC term, and no depth-of-discharge input. Fit: NREL reports 1.4% RMS capacity error across nine conditions from 0 to 55 °C; Schimpe et al. report end-of-test error below 1% of original capacity on a stationary-storage profile with a maximum relative error below 21%. Data: Sandia's multi-year study shows how widely results spread with chemistry and conditions — LFP cells cycled 2,500–9,000 equivalent full cycles, most still above 80% capacity when the study closed, against 250–1,500 for NCA and 200–2,500 for NMC, and round-trip efficiency varied up to 10% among fresh cells and fell over 5% with age. Limits: NREL/NLR's own open-source documentation says its models predict expected life, not warranty life, are trained on 1–3 years of accelerated data extrapolated to 10+ years, should be treated as qualitative, do not cover abuse, damage or pack imbalance, and offers as a rule of thumb that pack life is 20–30% below cell life. The explanatory note to a 2026 Chinese consultation draft adds the system-level gap: by participating units' analysis of operating-station data, inconsistency between cells can cost a cluster up to 5–15% of usable energy in the middle and late operating period, which is why its authors refused to equate cluster SOH with the worst cell. Verification: the PNNL/Sandia Reference Performance Test, IEC 62933-2-1 and the GB/T 36276 / 34131 / 36558 / 36549 series are the hooks for checking a model against the plant. This guide reads the papers, documentation and standard records; it quotes no plant dataset, warranty curve or augmentation schedule.

1. What a semi-empirical model contains

CNWattHub reading: the two published models disagree on a point that matters for a model review — whether depth of discharge belongs in the cycle term. Schimpe et al. leave it out for their LFP cell and use throughput. A reviewer should ask whether the vendor's model uses depth of discharge, throughput or both and whether the choice was fitted to the cell actually procured, because the two give different answers for shallow, frequent cycling.

2. The data behind the models: what laboratory aging shows

• Chemistry gap. Under the examined conditions LFP cells cycled 2,500–9,000 equivalent full cycles, NCA 250–1,500 and NMC 200–2,500; most LFP cells had not reached 80% capacity when the study closed, so the LFP range is observed cycling rather than a measured life to 80%.
• Condition sensitivity. Even inside manufacturer-specified ranges, cycles to 80% varied by thousands with conditions; the common LFP/NMC test matrix covered 0–100%, 20–80% and 40–60% windows at 15, 25 and 35 °C and discharge rates from 0.5C to 3C, with NCA excluded from 3C discharge; the matrix is not a full combination of every window, temperature and rate.
• Efficiency is not constant. Round-trip efficiency varied up to 10% among fresh cells, fell more than 5% with aging, and decreased consistently with discharge rate; LFP cells generally had the higher efficiency.
• BLAST-Lite's extrapolation caveat. The BLAST-Lite documentation says its models are trained on data collected over 1–3 years and extrapolated to predict real-world lifetimes of 10+ years, and should be treated as qualitative; it states this for its own models, not for every degradation model.

CNWattHub reading: a model that shows a single efficiency and a single cycle-life number for the project's life contradicts the laboratory evidence on two counts — efficiency falls with age and with rate, and cycle life is a range measured in thousands. The review question is not "what is the cycle life" but "at which temperature, window and rate was it measured, and does the site's duty match".

3. From cell to pack to plant

CNWattHub reading: the 5–15% cluster figure comes from an explanatory note to a consultation draft and is attributed to the participants' own station data, so it is an indication, not a standard's requirement; its use here is to show that a project model which converts a cell curve into a plant curve with no inconsistency term is omitting a loss the Chinese drafting group found large enough to argue about.

4. The verification hooks

• Reference Performance Test. The PNNL/Sandia protocol defines it as tests at the beginning of life to set the baseline and at periodic intervals to determine degradation, and requires stored-energy capacity tests on the same system to stay consistent so that degradation can be tracked.
• IEC 62933-2-1:2017. Unit parameters and testing methods for electrical energy storage systems at system level (TC 120; devices and technologies out of scope).
• Chinese standards. GB/T 36276-2023 (storage lithium-ion batteries), GB/T 34131-2023 (battery management systems), GB/T 36558-2023 (general technical requirements for electrochemical storage systems) and GB/T 36549-2018 (operating indicators and evaluation of electrochemical storage stations) — all current recommended national standards under the China Electricity Council; their texts were not read for this guide.

CNWattHub reading: a degradation model is only auditable if the plant has a baseline and a repeatable periodic test; the Reference Performance Test is the mechanism, and the warranty's test conditions should match it. For the Chinese standards named here, the question to put to an EPC or BMS supplier is which of them it cites for capacity, SOH reporting and station indicators, and on which clauses; what each requires of SOH estimation is not described here because the texts were not read.

5. An assessment sequence built from the sources

1. Structure. Separate calendar and cycle terms; temperature, SOC, throughput and rate as inputs; whether depth of discharge is used and why.
2. Fit evidence. RMS or end-of-test error on data not used for fitting; conditions covered.
3. Cell match. The model fitted to the procured cell's chemistry and format; observed cycling ranges by chemistry as a sanity check.
4. Duty match. Site temperature, SOC window, cycle depth and rate against the test matrix.
5. Efficiency. Modelled as a function of age and rate, not a constant.
6. Extrapolation. Years of underlying data versus project life; 'expected' versus 'warranty' life.
7. Pack and cluster. Cell-to-pack derating and an inconsistency term, with the basis stated.
8. Field test. Reference Performance Test schedule and consistency; standards cited by the supplier.

Scope limits

This guide reads NREL's 2017 life-prediction preprint, Schimpe et al. (2018), Preger et al. (2020), the BLAST-Lite README, the PNNL/Sandia 2016 protocol, the IEC 62933-2-1:2017 abstract, the records of GB/T 36276-2023, GB/T 34131-2023, GB/T 36558-2023 and GB/T 36549-2018 (texts not read), and the April 2026 explanatory note to a CNESA consultation draft (not an issued standard). It quotes no plant dataset, warranty curve, augmentation schedule, vendor model or cost figure. Not engineering, warranty or investment advice.