Calafat — Battery Cycle Deep Dive
7 Aug 2026 (S13, D69) · basis: as-drawn Eren config (7,3083 MWp · 16 MWh / 4 MW · €1.050k + all-in EPC),
verified Studio engine, post-2021 price basis · 6-agent adversarial workflow (every headline independently attacked;
two refuted claims corrected below) · INTERNAL — numbers here never go outward.
One-paragraph verdict: the battery earns by buying night-grid power and surplus midday solar,
selling it into the evening peak — the grid is its main fuel (48–70% of charge), and permission to grid-charge is
worth +2,87 IRR points. Cycling harder buys nothing: the policy surface is a plateau and the economic ceiling is
~495 cycles/yr. Once battery wear is priced per cycle, the optimum shifts slightly gentler than today
(spread bar €20–30 → ~335–395 cyc/yr, +0,13 pts). The standing SOH caveat is closed: central effect −0,02 pts
(noise); the real exposure is the un-warranted tail (≈ −1,5 pts) — which the written SOH table + a
≥460 cycles/yr allowance buys off. Ask paper A.12 updated accordingly.
1 · A day in the battery's life (pure observation, bit-exact vs the engine)
- Cycles per price-year: 2021: 373,7 · 2023: 441,8 · 2024: 437,2 · 2025-26: 424,4 — basis avg 417,6/yr = 1,14/day.
A metronome, not bursts: 71–81% of days sit in the 1–1,5 band; the 2-cycle daily cap is touched 0–1 day a year; idle days 0–2.
- Charge fuel: grid 48–70% of charged MWh (avg €67–81/MWh + €10 network charge) · free clipped solar 18–24% at €0 ·
diverted exportable solar rising 7% → 34% across the years as midday prices collapse (2025-26 divert cost only €30,4/MWh).
- Realized margin per cell-MWh: €78 (2021) · €83 (2023) · €112 (2024) · €108 (2025-26) — the average trade clears the €15
threshold 5–7 times over.
- The clock (wall time): buys 02:00–07:00 (night grid) + a 12:00–15:00 solar shoulder; sells 18:00–22:00 (67–72% of all
discharge) + a morning peak 07:00–09:00.
- Seasonality: cycle count is nearly flat; arbitrage € follows the price regime (2021 was Q4-loaded, 2024 peaked in July at €139k/mo).
- Curtailment is 53–70% self-hedged: a curtailed hour converts blocked exports into free charge; net cost of the 8% assumption
is only €17–23k/yr. (Cut days out-earn clean days in 2023/24/25-26 — not in 2021.)
2 · What actually limits cycles (a refuted claim, corrected)
The anatomy pass first blamed PCS power. The adversarial verifier refuted it by experiment: upsizing the PCS 4→5 MW adds
only +5–7 cycles/yr (+1–2%), while removing the €15 spread threshold adds +59–98 cycles/yr (471–511/yr). Corrected statement:
the spread threshold limits cycle volume; PCS width limits profitable volume at that threshold — the extra
sub-threshold cycles carry almost no money (arb +1–3%). The 2-cycle daily cap and gate room are irrelevant in practice.
Economic ceiling ≈ 495 cyc/yr (threshold at zero); anything past that requires trading at a loss (forced runs reach ~728).
3 · Policy ladders — "anything approved" measured (wear-blind engine)
| Policy | Equity IRR | cyc/yr (/day) | Arbitrage €/yr | Grid bought MWh |
|---|
| PV-only (grid charging OFF) | 8,96% | 237 (0,65) | 406.450 | 0 |
| maxCyc 1 (one cycle/day cap) | 11,20% | 350 (0,96) | 565.951 | 3.161 |
| DEFAULT — thr €15 · maxCyc 2 · grid ON | 11,83% | 418 (1,14) | 610.070 | 4.040 |
| Warranty-max ≤430 — thr €13 · maxCyc 3 | 11,84% | 429 (1,17) | 612.646 | 4.202 |
| Unconstrained max — thr ~€11,5 · maxCyc 3 | 11,84% | 437 (1,20) | 614.568 | 4.343 |
| Threshold zero (every profitable trade) | 11,74% | 495 (1,36) | 619.346 | — |
- Grid-charging permission = +2,87 pts (8,96% → 11,83%). Nearly all battery value beyond PV-shifting comes from buying cheap
night hours. "Fill from PVs" alone is a 9%-IRR battery.
- The top is a plateau: every sane policy (thr 10–15, maxCyc ≥ 1,5) is within ~0,04 pts. Marginal cycles are near-worthless:
net of the €10/MWh import network charge, the 418→434 step earns ~€71 per cycle-year and thr 12→0 is net-negative.
- Sensitivities worth money elsewhere: round-trip efficiency +0,21 pts per point (0,875→0,90 = +0,53 pts — a PCS/battery
spec item for the TEKLİF); import network charge ≈ −0,98 pts per +€10/MWh (gtar €10 assumed — verify DEO's actual import tariff);
payback slips 8,5→9,5 yr at gtar €20+.
4 · Wear priced in — the SOH verdict (and a second refuted claim)
The engine ages the battery by calendar only (vendor-style table, ~1,56%/yr, replacement year 15 at €2,16M); cycles never
enter wear, and an aged battery still dispatches a full 16 MWh. A throughput-indexed model was built and externally calibrated
(central: 7.000 full cycles → −20% SOH, +0,75%/yr calendar; band 6.000–8.000 validated against published LFP data; site duty
0,25C is gentler than test conditions).
- The standing caveat ("1,14–1,18 cyc/day vs ~1/day documented") is CLOSED — central net effect −0,02 pts at 418 cyc/yr, −0,05
at 430: noise. In-life extra wear and later-than-assumed replacement nearly cancel.
- The tail is the real exposure: at 6.000-cycle life + 1,5%/yr calendar fade the net hit is ≈ −1,4/−1,5 pts (replacement
year ~11, twice in 30 years). This tail — not the central case — is what the written SOH table + warranty condition buys off,
and it independently justifies the €60k extended-warranty price (D62).
- Refuted-and-corrected optimum: the producer placed the optimum at today's 418–434. The verifier showed that under the
throughput-wear model the optimum is gentler: spread bar €20–30 → ~335–395 cyc/yr ≈ 11,94–11,95% vs 11,81% at today's
settings (+0,13 pts) — marginal cycles net ~€71 each while costing ~0,003 pts of battery life each; the wear-vs-arbitrage
crossover sits near ~360 cyc/yr. Direction: trade a little less, not more. (A thr operating setpoint is free to change
at any time — software, not hardware.)
5 · The warranty demand (feeds ask paper A.12 — updated)
- ≥ 460 cycles/yr allowance — a 430 allowance would already be breached in 2023/24-price years (437–442 cycles at default
dispatch). 460 gives operating room; its actuarial cost to the vendor is small (central time-to-70% moves 15,4 → 14,5 yr).
- The vendor's own SOH table written in as the warranted floor at that allowance (central-model reference points at
actual 418–434 operation: year 5 ≈ 90% · year 10 ≈ 80% · year 15 ≈ 70%).
- Pin the cycle-counting denominator: BOL nameplate (16 MWh). Metered against actual (shrunken) capacity, an aged battery
logs more cycles for the same work — a hidden allowance-eater.
- Fallback if the vendor holds at 430: cap dispatch via the spread threshold (thr ≈ €12–13) — costs ≈ €1k/yr net, noise
against voiding cover on a €2,16M replacement. Never cap via a hard daily limit (maxCyc 1,18 costs −0,26 pts).
6 · What changes
- Ask paper A.12 updated before sending: allowance ≥430 → ≥460 + BOL-nameplate counting added (regenerated 7 Aug).
- Operating policy note: once real, run the plant at a €20–30 spread bar under a throughput-wear view (+0,13 pts vs default);
revisit when the vendor's written SOH table arrives.
- Studio backlog: optional cycle-indexed SOH knob (central calibration above); verify DEO's actual import network tariff
(model carries €10/MWh; each +€10 ≈ −1 pt); round-trip efficiency is a spec lever worth +0,2 pts/point.