Estimating state of health from operational data, without a test cycle

State of health is quoted as a percentage and can mean either of two different things. Capacity fade, the usable capacity now as a fraction of rated capacity. Or resistance growth, the internal resistance now relative to when new, which determines available power rather than available energy.

They diverge. A cell can hold most of its capacity while its resistance has risen enough that it can no longer deliver rated power, and a duty that is power-limited cares about the second while a duty that is energy-limited cares about the first.

Which means a state of health figure without its definition attached is close to useless, and a warranty or a report that quotes one without saying which is inviting a disagreement later.

It is also not directly measurable in either sense. It is estimated from quantities that respond to it, which makes the method behind the number the thing worth examining.

The clean way to measure capacity is a full controlled discharge from a defined starting condition at a defined rate and temperature. On a grid-connected battery that means taking the asset out of service, foregoing revenue, and putting the pack through a cycle it would not otherwise do.

Owners are reluctant, and reasonably so. The test consumes cycle life to measure cycle life, and on an asset with a market obligation it is genuinely expensive.

So the estimate has to be built from operating data: partial cycles, at whatever rate the market called for, at whatever temperature the day provided. That is the harder problem, and it is the one an owner is left with.

The workable approaches all rely on finding windows in the operating record that are clean enough to be informative and then correcting them.

Coulomb counting across a window between two well-determined states of charge gives a capacity estimate directly, and its accuracy depends entirely on how well those two endpoints are known. Endpoints determined near the flat part of the voltage curve are poorly determined, which is why window selection matters more than the arithmetic.

Voltage-based methods use features of the relationship between voltage and charge that shift in characteristic ways as a cell ages. They need a reasonably wide window and a low enough rate that the measurement is not dominated by resistance effects, and market duties sometimes provide those and sometimes do not.

Model-based estimation runs an electrochemical or equivalent-circuit model alongside the pack and adjusts its parameters to match observed behaviour. It uses every window rather than only the clean ones, at the cost of depending on the model being right.

In all three the honest output includes how much data went into it and how well conditioned that data was. An estimate from a month of shallow high-rate cycling is weaker than one from a month that happened to include a few deep slow ones, and reporting them identically hides that.

Internal resistance is easier to estimate from ordinary operation than capacity is, because every current step is a small experiment. The voltage response to a change in current, corrected for temperature and state of charge, gives resistance without any special test.

It frequently moves before usable capacity does, particularly on duties involving high rates, which makes it the earlier indicator on many grid assets and the more useful one for maintenance decisions.

It is also strongly temperature dependent, more so than capacity, so an uncorrected resistance trend is largely a trend in ambient conditions. Resistance estimates compared without stating the temperature and state of charge they were taken at are not comparable at all.

Its relationship to warranty is worth noting: most battery warranties are written on capacity retention, so resistance growth can be operationally important and contractually invisible. Both should be tracked for that reason alone.

Temperature affects both apparent capacity and resistance substantially, and a cold pack looks degraded. Rate affects apparent capacity, and a high-rate discharge delivers less than a slow one from the same cell. State of charge affects resistance measurement.

Correcting for all three is what turns a set of estimates taken across months of varying operation into a trend that means something. Without it, seasonal variation dominates and the resulting curve is mostly weather, which is the same error that ruins uncorrected gas turbine performance trends.

The correction needs a reference, and the reference is a characterisation taken while the asset was new. The pattern repeats across every asset class: the baseline is cheap at commissioning and cannot be recreated afterwards. Where none exists, the honest position is that absolute degradation cannot be recovered and only relative trends from now are available.

A pack state of health is an aggregate across a very large population of cells, and how it is aggregated depends on the topology, the balancing, the limits applied and the algorithm the battery management system uses. A small number of cells substantially degraded barely moves it, in exactly the way a single degraded cell barely moves an electrolyser stack voltage.

The distribution decides the maintenance action. A population that has aged uniformly is an asset approaching a planned replacement or augmentation. A population with outliers is a candidate for module-level intervention, and the two have very different costs.

It also decides the risk picture, since cells in poor condition are the population that matters for the slow problem described in thermal runaway: what is predictable and what is only detectable.

Whether the distribution is visible at all depends on what the battery management system exposes and at what granularity, and that is a procurement decision made long before anyone wants the data. It is worth specifying explicitly rather than discovering later.

The definition: capacity fade or resistance growth. The conditions it was normalised to: temperature, rate and state of charge window. The method and the amount of data behind it. The uncertainty. The date. And whether it is a pack aggregate or a distribution, with the spread if it is the latter.

Six items, and a state of health figure quoted without them is not auditable, which matters when the figure is feeding a warranty claim, a reserve calculation or a valuation.

It is the same discipline as any other estimated quantity on a plant, and the reason it keeps recurring is that estimated quantities are the ones commercial decisions get made on.