Parallel connection

What it is (precise)

Connect two elements terminal-to-terminal — positive to positive, negative to negative — and they share one pair of electrical nodes. Kirchhoff's voltage law then forces the same voltage across both, and Kirchhoff's current law makes the total current the sum of the branch currents. Charge capability adds the same way: parallel amp-hours sum.

Take the standard rack — 416 series LFP cells at 314 Ah, about 418 kWh (the series-connection entry derives it). A second identical rack in parallel gives 628 Ah at the same voltage, about 836 kWh. Series connection is the exact mirror: voltage adds while amp-hours stay constant, and only their product, energy, accumulates through both directions of aggregation.

The configuration is written in xPyS notation — parallel groups times cells in series, as in 1P104S. Parallel also aggregates differently from series in kind, not just in axis. A series string is weakest-cell limited: every cell carries the same current, so the worst cell caps the whole string. A parallel group is additive: a weak element subtracts only its own shortfall from the total.

And at rest, the shared nodes are self-correcting — any open-circuit-voltage difference between paralleled elements drives an equalizing current until they agree — whereas a series string never equalizes itself and needs the BMS to balance it. That forgiveness at rest is real, but it is bought with the current-sharing and circulating-current behaviour below.

Where parallel appears in the hierarchy

At cell level, parallel groups are welded or wire-bonded inside the module — cylindrical formats bond many small cells into each parallel block, with the wire bonds themselves sized to act as cell-level fuses so a failed cell is isolated without dropping the string.

Modern grid-scale design has largely moved the parallelism out of the module: 5 MWh-class containers run 1P104S modules — one large prismatic LFP cell per series position — because a 314 Ah cell already carries the whole string current on its own. The P in the vendor's notation is worth checking anyway, since the same parallel count made at cell, module or rack level produces different balancing and failure behaviour.

The dominant parallel connection in a utility-scale plant is racks onto a shared DC bus. Each rack is an independently protected series string — its own contactor, fuse, pre-charge circuit and rack BMS — and several of them parallel inside the enclosure to feed the PCS; in the DC fuse hierarchy, racks in parallel form a section and sections in parallel form a container.

The parallel count is also what builds DC fault current: a battery rack typically supplies a fault current of 1 to 12 kA, while a combined system can reach 250 kA or more once racks and sections are paralleled. DC protective devices are therefore rated against the paralleled bus they sit on, not against the single string behind them.

Current sharing — and why it skews

Because the voltage across paralleled elements is identical, the current divides in inverse proportion to each branch's impedance. Identical racks with identical connections would share equally; real racks do not, because the loop impedances differ — cable run lengths to the bus, joint and contactor resistances, temperature, and age all enter. The lower-impedance branch takes more than its share of every charge and discharge, which means more throughput, more I²R heat, and a duty the fleet-average numbers never show.

Skewed sharing compounds over years. The rack doing extra work cycles harder and runs warmer, so the fleet ages unevenly — and the shared bus hides it, because bus voltage says nothing about which rack carried the current. The per-rack current logs in the BMS are the evidence, and they are worth asking for whenever a paralleled system underperforms a capacity test: a healthy average can conceal one rack limiting early and its neighbours quietly covering the shortfall.

The two failure directions differ sharply. An element that fails open — a blown rack fuse, an opened contactor — simply removes its amp-hours, and its share of the current redistributes onto the remaining branches. A short is the dangerous direction: every paralleled neighbour feeds it at once, which is precisely the job rack and section fuses exist to interrupt, and at cell level why parallel-block designs build fusible links into the cell contacting system.

Circulating current and the paralleling event

Connect two strings whose open-circuit voltages differ and current circulates between them — the voltage difference divided by the loop impedance — with no load connected at all. The loop is built from busbars and joints deliberately sized for full operating current, so its resistance is low, and even a few volts of mismatch drives an equalizing current on the scale of a real dispatch.

The flat LFP voltage curve makes this worse than it looks: on the plateau around 3.2-3.3 V per cell, a large state-of-charge difference shows up as a small per-cell voltage offset — but multiplied across 416 series positions, small offsets become tens of volts at the string terminals.

This is why closing a rack onto an energized bus is a controlled event, not a switch flip. The rack BMS runs a pre-charge sequence that equalizes voltage across the main contactors before they close, and the system-level BMS arbitrates which strings may close onto the shared bus at all — a rack returning from maintenance at a different state of charge is brought into a voltage window first.

Augmentation is the fleet-scale version of the same problem: new full-capacity racks paralleled with faded ones will not share equally, and how the design bounds the new racks' current share is a question to put to the integrator before the racks arrive, not after.

Common pitfalls

The arithmetic trap: cell amp-hours times parallel count does not translate to deliverable megawatt-hours. Between the parallel groups and the revenue meter sit the series architecture, the SOC window, conversion and auxiliary losses — a nameplate assembled from cell arithmetic overstates what the meter will record. Parallel addition is exact for amp-hours at the paralleled node; it says nothing about energy at any boundary further downstream.

The modelling traps are quieter. Treating the system rating as per-rack rating times rack count assumes equal sharing that real impedances do not deliver, and it fails soonest in mixed-age fleets. Sizing DC protection against single-rack fault current ignores that the paralleled bus can deliver an order of magnitude more into the same fault. And comparing vendor xPyS notations without asking where the parallel connection is physically made — inside the module, at the rack, on the bus — compares configurations that fail, balance and fuse differently while wearing the same label.