CHENLUNBATTERY SIMULATION
← Selected work案例研究
CASE 02 · AGEING & LIFETIME PREDICTION · S-05 / S-07

Knee onset predicted from the first 300 cycles用前 300 次循环预测衰减拐点

A storage integrator was qualifying a second cell supplier and had to commit to a warranty long before the supplier's test cells would reach end of life. We built a degradation-mode model that predicts where capacity fade turns non-linear using only early data, then tested it on 32 cells it had never seen.

一家储能集成商在导入第二家电芯供应商时，必须在对方测试电芯走完寿命之前就签下质保承诺。我们建立了基于衰减机理的模型，只用早期数据预测容量衰减何时转为非线性，并在 32 颗模型从未见过的电芯上检验。

142 cycles次
KNEE-ONSET RMSE, 32 HELD-OUT CELLS32 颗留出电芯拐点 RMSE
30 / 32
PREDICTED WITHIN ±250 CYCLES误差在 ±250 次以内
4.4 %
MEAN ABSOLUTE ERROR平均绝对百分比误差
300
CYCLES OF DATA NEEDED PER CELL每颗电芯所需循环数据

01The problem问题

LFP cells fade almost linearly for a long time, then bend sharply downward. That bend, the knee, is what sets warranty risk, and it cannot be seen in early data by eye. On the supplier's previous cell generation, extrapolating the linear phase had over-predicted life by 35–60 %.

LFP 电芯在很长一段时间内几乎线性衰减，随后突然加速下滑。这个转折点——拐点——决定了质保风险，而在早期数据中肉眼无法看出。对该供应商上一代电芯，用线性段外推曾把寿命高估了 35–60 %。

02What we did我们做了什么

Split the data honestly严格划分数据

96 cells across six conditions (25 / 35 / 45 °C × 0.5C / 1C, 100 % DoD). 64 cells for calibration, 32 held out, split by cell and stratified by condition, so no held-out cell shares data with training.

96 颗电芯覆盖六种工况（25 / 35 / 45 °C × 0.5C / 1C，100 % DoD）。64 颗用于标定，32 颗留出；按电芯划分并按工况分层，留出电芯的数据完全不进入训练。

Separate the degradation modes拆分衰减模式

At each reference performance test we fitted dV/dQ curves to separate loss of lithium inventory (LLI) from loss of active material on each electrode (LAM).

在每次 RPT 中拟合 dV/dQ 曲线，把活性锂损失（LLI）与正、负极活性材料损失（LAM）分开。

Model the mechanism behind the knee对拐点背后的机理建模

SEI growth drives LLI; temperature- and rate-dependent LAM erodes the graphite's capacity margin. When the anode no longer has room for the lithium inventory, plating begins and fade accelerates. Bayesian calibration gives each cell its own uncertainty band.

SEI 生长导致 LLI；与温度、倍率相关的负极 LAM 不断侵蚀石墨的容量余量。一旦负极容纳不下可用锂，便开始析锂，衰减随之加速。贝叶斯标定为每颗电芯给出各自的不确定度区间。

Predict from early data only只用早期数据预测

For each held-out cell the model saw only the first 300 cycles and three RPTs, then predicted the knee onset and the full fade curve.

对每颗留出电芯，模型只使用前 300 次循环与三次 RPT，预测拐点位置与完整衰减曲线。

03Results结果

Measured capacity fade of four held-out cells with model predictions, uncertainty bands and predicted knee markers.
FIG 1 · Four representative held-out cells. The shaded region is all the model saw; lines and bands are predictions; ▼ marks the predicted knee onset.图 1 · 四颗代表性留出电芯。灰色区域是模型可见的全部数据；曲线与色带为预测值；▼ 为预测拐点。
Predicted versus measured knee onset for 32 held-out cells.
FIG 2 · Predicted vs measured knee onset. Band = ±250 cycles.图 2 · 拐点预测值与实测值。色带为 ±250 次。
Stacked capacity loss by degradation mode at 45 °C.
FIG 3 · At 45 °C, anode LAM, not SEI, accelerates into the knee.图 3 · 45 °C 下，推动拐点的是负极 LAM，而不是 SEI。
  • Knee onset RMSE of 142 cycles on 32 held-out cells whose knees fell between 1,900 and 3,600 cycles. 30 of 32 predictions were within ±250 cycles.
  • 拐点预测 RMSE 为 142 次循环，32 颗留出电芯的拐点分布在 1,900–3,600 次之间；32 颗中有 30 颗误差在 ±250 次以内。
  • The mechanism changed the operating advice. Anode LAM at high temperature drives the knee. Lowering the container set-point from 35 °C to 30 °C in hot-climate sites is projected to add 18 % more cycles before the knee.
  • 机理结论改变了运行建议。高温下的负极 LAM 是拐点的主因。将炎热地区储能柜的温控设定从 35 °C 降至 30 °C，预计拐点前可多出 18 % 的循环次数。
  • A knee-based acceptance test. The supply contract now includes a 300-cycle screening test with the model's P10 prediction as the pass criterion.
  • 以拐点为依据的验收测试。供货合同中加入了 300 次循环的筛选测试，以模型的 P10 预测值作为合格判据。

04Limits of validity适用边界

LFP | graphite cells from this supplier, 25–45 °C, 0.5–1C, 100 % DoD. Partial-DoD cycling and calendar ageing at rest are estimated but not yet validated: only nine months of calendar data existed at handover. The pipeline re-runs as new RPT data arrive.

适用于该供应商的 LFP | 石墨电芯，25–45 °C、0.5–1C、100 % DoD。部分放电深度循环与静置日历老化目前为估算值、尚未验证——交付时日历老化数据仅有九个月。新的 RPT 数据到达后流水线会自动重新计算。

05What the client received交付内容

Model and calibration pipeline that re-runs on new RPT data
可随新 RPT 数据自动重跑的模型与标定流水线
Per-cell predictions with P10 / P90 bounds
每颗电芯的预测值及 P10 / P90 区间
Warranty life table for the site duty cycle
针对站点工况的质保寿命表
Handover workshop for the reliability team
面向可靠性团队的交接培训
NEXT · CASE 03 Four data systems, one question box →四个数据系统，一个提问框 →