Arrhenius Capacitor Aging: What It Really Means for Inverter RUL

Key facts

Electrolytic DC-link capacitors are the second-most failure-prone power component in PV inverters, after IGBT modules.
The Arrhenius rule of thumb — lifetime halves for every ~10 °C above the rated core temperature — derives from an exponential dependence of electrolyte loss on temperature.
Voltage stress and ripple current modify the rule: real-world lifetime is governed by combined thermal-electrical aging, not temperature alone.
Capacitance drops and ESR rises as the capacitor ages — both observable indirectly through DC-link ripple in standard SCADA data.
A calibrated Arrhenius + Power-Law (Eyring-style) model paired with ripple-current trending produces RUL estimates with ±20% confidence against observed field failures.
In a 5 MW central inverter, a DC-link capacitor failure typically means a 1-3 day forced outage and €4-15k in unsold energy — dominantly avoidable with weeks-ahead RUL.

Why DC-link capacitors matter more than they look

In a utility-scale PV inverter, the DC-link capacitor bank sits between the array and the IGBT bridge. Its job is unglamorous but critical: hold the DC bus voltage stable, absorb the high-frequency ripple from the switching bridge, and provide a stiff voltage reference during cloud-induced power transients. Without it, the bridge cannot operate.

In most central inverters the bank is built from aluminium electrolytic capacitors — chosen for their unrivalled volumetric capacitance and cost per joule. The trade-off is that they are the only component in the inverter built around a liquid electrolyte that slowly evaporates through a rubber seal. That liquid is what ages, and it ages primarily as a function of temperature.

Reliability data from Fraunhofer ISE and OEM warranty databases consistently rank DC-link capacitors as the second-most failure-prone power component in PV service, behind IGBT modules and ahead of gate drivers, fans and contactors. A single bank failure in a 5 MW block typically forces a 1-3 day outage while the replacement is sourced and installed — at current Iberian spot prices that means €4-15k of unsold energy, dwarfing the cost of the part.

The Arrhenius equation in capacitor reliability

Svante Arrhenius described chemical reaction rate as an exponential function of temperature:

k = A · exp(−Ea / (R · T))

where k is the reaction rate, A a pre-exponential constant,Ea the activation energy of the reaction, R the gas constant, and T the absolute temperature.

In an aluminium electrolytic capacitor, the dominant aging mechanism — electrolyte evaporation through the rubber bung — follows exactly this form. The consequence is the well-known rule of thumb often quoted in datasheets:

L = L_rated · 2^((T_rated − T_core) / 10)

For every 10 K of core temperature reduction below the rated value, life approximately doubles. For every 10 K of excess, it approximately halves. The factor is an empirical fit to the Arrhenius exponential and works well as a first approximation — between roughly 40 °C and 105 °C core temperature, which is the useful envelope for PV inverter banks.

Why the rule of thumb is not enough

Two practical failure modes break the simple temperature-only model:

Ripple-current self-heating

The capacitor's own ESR turns ripple current into heat dissipated inside the core. Two capacitors at identical heatsink temperature but different ripple-current loads will have different core temperatures — and therefore different ages — even though the inverter cabinet looks the same on the outside.

Worse, this is a positive-feedback loop. As the capacitor ages, electrolyte loss raises ESR. Higher ESR means more ripple heating, which raises core temperature, which accelerates electrolyte loss. The end of life is non-linear, and a temperature- only Arrhenius model will be optimistic about the late-stage degradation rate.

Voltage stress

Field-strength effects in the dielectric add a second exponential term. A capacitor running at 90% of rated voltage ages materially faster than one at 60%, at the same temperature, because the oxide layer hosts more parasitic reactions. This is captured by an Eyring-type extension:

L = L_rated · 2^((T_rated − T_core) / 10) · (V_rated / V_op)^n

with n typically in the range 3-5 for aluminium electrolytics in service voltage ranges. The combined thermal-electrical model is what InverterAI uses internally; the temperature-only version is reserved for first-cut envelope estimates.

Observable signatures of an aging capacitor

Direct capacitance and ESR measurement requires lab equipment and a disconnected unit. Online prognostics has to extract those quantities indirectly from SCADA data. Three signatures are reliable in practice:

DC-link voltage ripple amplitude at the line and switching frequencies grows as capacitance drops. A 15-20% rise from the commissioning baseline is a strong indicator that the bank is past mid-life.
Harmonic content of the AC-side current changes subtly as DC-link stiffness degrades. The 5th, 7th and 11th harmonic signatures track this with high fidelity in the InverterAI feature pipeline.
Audible 100/120 Hz hum from the inverter cabinet, reported by site technicians, is the late-stage symptom. By then RUL is on the order of weeks, not months — InverterAI's job is to predict the same condition months earlier.

From SCADA to RUL forecast: the pipeline

The InverterAI pipeline for capacitor RUL combines four ingredients:

Core-temperature inference from heatsink and ambient telemetry, using a calibrated thermal-network model analogous to the one used for IGBTs but tuned to the bank's thermal mass and ventilation path.
Estimated ripple-current trace from AC-side current harmonics and switching-frequency telemetry. Where direct DC-link current sensors exist (newer inverters), they short-circuit this estimation step.
Calibrated Arrhenius + Power-Law model with constants tuned per capacitor family and OEM. These come from datasheet curves cross-checked against fleet-observed failure timing.
Bayesian update as new SCADA windows arrive: the RUL posterior tightens (or, occasionally, shifts) when the observed ripple trajectory disagrees with the prior.

The output is the same shape as for IGBTs — a RUL distribution with a 90% confidence band — and feeds the same prioritised maintenance queue. Capacitor and IGBT predictions share the same ranking; whichever fails first dictates the intervention.

What predictive capacitor RUL changes operationally

Capacitors are a particularly interesting case because the failure is binary at the inverter level: when the bank goes, the inverter stops. The economics of advance warning are therefore stark:

Spares arrive before they are needed. A high-power capacitor bank for a central inverter can have 4-12 week lead times from the OEM, longer for legacy platforms. Six-week RUL advance warning is the difference between a planned swap and an emergency-priced expedite.
Crews are scheduled, not scrambled. Bank replacements take 4-8 hours of work in the cabinet. Doing them on a planned low-irradiance day is an order of magnitude cheaper than doing them on a peak-production day with emergency-rate technicians.
The replacement itself can be optimised. A documented degradation trail justifies upgrading to a higher-rated or different-chemistry capacitor at replacement time — pushing the next failure further out.

Honest caveats

Production defects bypass the model. Manufacturing variability, shipping damage and storage-induced electrolyte drift produce failures that look nothing like wear-out. They show up as wide error bars in the early-life RUL prediction; the model handles them by widening the confidence band rather than pretending to forecast them.
OEM datasheet curves are conservative. Real fleet data typically shows longer life than the datasheet predicts at the same operating point. The InverterAI model uses datasheet curves as a prior and updates with field observations — but until enough field data accumulates, predictions skew conservative.
Mixed-capacitor banks complicate the model. Some central inverters parallel different capacitor part numbers within the same bank. The aging prediction has to track each population separately, then aggregate to bank-level health — adding model complexity but not insurmountable.
The model predicts wear-out, not catastrophe. Internal short circuits from manufacturing defects, electrolyte leakage from seal failures, and physical damage are not in the Arrhenius envelope and never will be. They are caught (if at all) by SCADA anomaly detection, not by RUL prognostics.