Cell balancing

What it is (precise)

In a series string every cell carries the same current, so charge and discharge terminate on whichever cell hits a voltage limit first — the string charges until its highest cell is full and discharges until its lowest cell is empty. Any state-of-charge spread between cells is therefore energy the string cannot sweep: within one string, measured at the DC terminals, each percentage point of SOC spread strands roughly a percentage point of the swept window.

Balancing is the function that removes the spread; why cells diverge in the first place — self-discharge mismatch, temperature gradients, manufacturing spread — belongs to Cell imbalance. Parallel-connected cells are largely exempt: they share a voltage by construction and equalize through their interconnection, which makes balancing a series problem.

The work sits inside the BMS hierarchy. Module-level boards sample per-cell voltage to roughly ±2-5 mV every 100 ms to a few seconds, and the rack- or string-level controller runs the balancing logic against those readings, selecting cells whose voltage sits a threshold above the string's low cell.

The sensing basis matters for LFP: across the middle of the SOC range the open-circuit voltage sits on a plateau around 3.2-3.3 V and moves only tens of millivolts, so mid-band voltage differences say almost nothing about SOC differences. The balancer gets its clearest view near the top of charge, where the curve steepens — which is where balancing activity concentrates in practice, during and after charge and through idle periods.

Passive versus active

Passive balancing is a bleed resistor per sensing channel, switched across whichever cells run high; the excess charge is dissipated as heat. It works only downward — high cells are trickle-discharged toward the low ones, never the reverse — and the currents are modest: roughly 50-200 mA per cell in stationary designs, a fraction of a watt per channel at LFP cell voltages.

What it buys is simplicity: one resistor and one switch per channel, repeated across the several thousand cells of a modern 5 MWh enclosure, with essentially nothing that can fail energetically. Grid-scale storage uses it almost universally.

Active balancing moves charge instead of burning it — DC-DC or switched-capacitor circuits shuttle charge from high cells to their neighbours or to a module bus, recovering the energy and supporting harder currents in both directions. The cost is a power converter where the passive design had a resistor, multiplied by thousands of channels, with the parts count, failure modes and price that implies.

The economics rarely close in stationary storage: a passive balancer dissipates only the charge it removes, and the charge it removes is bounded by the spread being corrected — a few percentage points of SOC, occasionally, on the high cells of a healthy string. Recovering that sliver rarely pays for the converters, and stationary designs stay passive.

The arithmetic of small currents

The defining fact about balancing is how slow it is. Modern large-format LFP cells sit in the 280-314 Ah class; one percent of SOC on a 314 Ah cell is 3.14 Ah, and at 50-200 mA of bleed current that single point takes roughly 16 to 63 hours to move. As a rate it is a few ten-thousandths of 1C — about C/1600 at the strong end of the range. A string that has drifted three percent apart is therefore days of continuous bleeding away from alignment, and no firmware setting changes that: the current is fixed by the resistor.

The consequence is that balancing is housekeeping, not an intervention. It cannot correct a meaningful spread inside a market day, so its real job is to outrun the divergence rate — removing spread slightly faster than mismatched self-discharge and thermal gradients create it, which for a healthy string demands only a trickle. The corollary is diagnostic: when the balancer visibly cannot keep up, suspect a cell diverging faster than any resistor could counter, and go find it. A larger bleed current would only burn more energy tracking a cell that belongs on an inspection list.

How it shows up in specs, tests and operations

On a module or rack datasheet, balancing is one line: the balancing current per cell, sometimes with the method (passive) and the voltage-difference threshold that triggers it. In operating data it appears as the max-min cell voltage per rack — one of the most useful health traces a SCADA point list can carry.

Flat and small means the balancer is winning; a spread that re-opens after every rebalance points at an outlier cell. Questions worth putting to a vendor: the per-cell balancing current, when the algorithm engages and stops, and whether per-rack spread and balancing status are exposed to the operator rather than kept inside the BMS.

Around capacity tests the timing matters. Strings that have sat through transport, storage or a long outage arrive with spread — self-discharge is not uniform — and a capacity test run before the fleet has been charged full and allowed to balance measures that spread as missing energy.

Test procedures begin from a full charge partly for this reason: the top of charge is where LFP balancing completes. The same effect runs through warranty measurements — a rebalance between two tests shifts the measured number independently of any real fade, which is one reason test protocols fix the preconditioning rather than leaving it to operational chance.

Common pitfalls

The first trap is conflating balancing with SOC-estimate calibration. Both happen near the top of charge, because the steep end of the LFP curve is the only place per-cell voltage resolves small SOC differences — but one moves real charge between cells and the other re-anchors an estimate, and a system can need either without the other. A fleet-average SOC that reads wrong is a calibration issue; a string that clips early at both ends is a balance issue; the trace that separates them is per-cell voltage spread, not the SOC display.

The second is arguing balancing architecture from the wrong magnitude. The energy passive balancing wastes is bounded by the spread it removes, so "passive wastes energy" is true and negligible at the same time — at 50-200 mA the dissipation per channel is a fraction of a watt, and it flows only while a high cell is actually being bled.

Equally, active balancing is not an upgrade a struggling string needs: at these capacities even an order of magnitude more balancing current corrects drift on a timescale of hours, and persistent divergence is a cell-condition signal that no balancer of any architecture should be sized to hide.