For most of the last century an island power system meant a diesel station. Frequency was held by the kinetic energy stored in the generator rotors and, a few seconds later, by the governors that admitted more fuel. As photovoltaic arrays, wind turbines and battery storage displace diesel running hours, the stored kinetic energy falls, and the frequency moves further and faster after the same disturbance (Tielens and Van Hertem, 2016). Islands meet this problem earlier than continental grids. Their systems are small, a single generator or feeder can carry a large share of the load, and there is no neighbour to lean on.

Most inverters installed to date are grid-following. They measure the terminal voltage with a phase-locked loop and inject a current in step with it, which presupposes that something else is holding the voltage. A grid-forming inverter reverses the roles: it presents a voltage of controlled magnitude and frequency behind an impedance and lets the current be decided by the network (Rocabert et al., 2012). The idea is not new. Parallel inverters sharing load by frequency droop were described three decades ago (Chandorkar et al., 1993), and the microgrid concept was built around it. What has changed is the scale at which system operators now contemplate running with few or no synchronous machines on line (Matevosyan et al., 2019; Lasseter et al., 2020).

This article reviews the published work on grid-forming control from the standpoint of one question: what it does for frequency stability on a small island system where inertia is scarce. It describes the control families, summarises what analysis and simulation have established, and identifies the limits and the unanswered questions that matter to an island utility.

The article is a narrative review of the published literature. It draws on the journal papers that introduced each control concept, on later comparative and survey articles, on standard textbooks and on the interconnection standards and roadmap reports that frame current practice. No claim is made to an exhaustive or systematic search, and no new simulations or measurements are reported. Stability terms follow the classification of the IEEE and CIGRE task forces (Kundur et al., 2004) as extended to converter-dominated systems (Hatziargyriou et al., 2021) and adapted to microgrids (Farrokhabadi et al., 2020). The emphasis is on islanded alternating current systems of the size found on small and medium islands, in which diesel sets, batteries and renewable plant share a medium-voltage network.

In a system of synchronous machines, frequency is a physical state. The swing equation relates the rate of change of rotor speed to the imbalance between mechanical and electrical power, scaled by the inertia constant H, the stored kinetic energy at rated speed divided by the machine rating (Kundur, 1994). Immediately after the loss of a generator, the initial rate of change of frequency is proportional to the size of the power imbalance and inversely proportional to the total stored kinetic energy. The lowest point reached, the nadir, depends in addition on how quickly governors and other reserves respond. Halving the inertia doubles the initial rate of change and leaves primary control half the time to act before under-frequency load shedding begins.

Milano et al. (2018) argue that the difficulty in converter-dominated systems is deeper than a smaller H. A converter has no rotor, the energy in its DC-link capacitor is negligible on the time scale of a frequency event, and its output frequency is whatever the controller computes. Frequency ceases to be a shared physical quantity and becomes a signal that the controls must create and agree upon. Microgrids add features of their own: feeders are short and often resistive, so the textbook decoupling of active power from voltage and reactive power from angle holds only approximately, and voltage and frequency stability cannot be studied apart (Farrokhabadi et al., 2020).

The literature contains many named schemes, but they fall into a small number of families that differ in how the converter derives its frequency. Figure 1 groups them, and Table 1 compares their synchronising principle, their contribution to frequency containment and the reservation most often raised against each.

Comparative simulation studies agree on the main point. When grid-forming converters replace grid-following ones at a high converter share, the system remains stable at shares where it would otherwise fail, and the frequency after a load step is better contained (Lasseter et al., 2020). A grid-forming unit responds to a change in network angle within its own control bandwidth and without first having to detect a frequency deviation, so the first reaction to an imbalance is not delayed by measurement. Tayyebi et al. (2020) compared droop, virtual synchronous machine, dispatchable virtual oscillator and matching controls in a common test system with synchronous machines and reported that all four improved the standard frequency metrics relative to the all-machine case, with differences between them that were modest in normal operation.

The same body of work shows where the trouble lies. The differences between schemes emerge under stress, when a converter reaches its current limit or its DC source cannot follow the demanded power, and the fast converter controls can interact adversely with the slow turbine and governor dynamics of the machines that remain (Tayyebi et al., 2020). For an island that keeps diesel sets for security and energy adequacy, this mixed regime is the normal one, and it is less thoroughly studied than either extreme. The stability classification for microgrids accordingly treats converter control interactions and power-sharing instability as categories in their own right (Farrokhabadi et al., 2020).

Three constraints recur. The first is current. A synchronous machine can deliver several times its rated current for a short period, whereas the semiconductor switches of an inverter tolerate little more than rated current. During a fault or a large load step the inverter must limit its output, by saturating a current reference or by inserting a virtual impedance (Paquette and Divan, 2015), and while it does so it no longer behaves as the voltage source that the stability analysis assumed. Synchronism can be lost during the limited period, and protection designed around large fault currents from rotating plant may not operate as intended.

The second constraint is energy. Synthetic inertia and fast frequency support are deliveries of real energy, and they must come from a battery, from curtailed renewable output held in reserve, or from the DC link. A virtual inertia constant can be set to any value in software, but the response is available only while the source behind the converter has headroom in both directions. The third is the network itself. On resistive low-voltage and medium-voltage feeders the active and reactive droop loops are coupled, and virtual impedance or transformed droop laws are needed to restore accurate sharing (Guerrero et al., 2011; Rocabert et al., 2012).

Standards have not yet caught up with these questions. IEEE Std 1547 is written for distributed resources that connect to an area system operated by others (IEEE, 2018), and IEEE Std 2800 sets requirements for inverter-based resources on transmission systems (IEEE, 2022). Neither defines in testable terms what a grid-forming unit on an isolated system must do, a gap noted in the research roadmap prepared by Lin et al. (2020).

Several gaps follow from the foregoing. There is no agreed performance specification for grid-forming behaviour that an island utility could write into a tender and verify at commissioning; the term currently covers devices with quite different responses. Planning tools are a second gap. Phasor models that neglect network and inner-loop dynamics can miss the control interactions described above, while electromagnetic transient models are costly to build and depend on vendor data that are seldom public. Guidance on which level of detail is sufficient for which study remains thin.

Third, most of the evidence is analytical or simulated. Operating islands that run for long periods on inverters alone exist, but detailed disturbance records from them rarely reach the journals, so the models have been checked against field behaviour less often than their number would suggest. Fourth, the literature on systems with units from several manufacturers, each with a proprietary control, is small, although this is the situation every utility will face after its second procurement. Black start and re-synchronisation of diesel sets to an inverter-formed network also deserve more systematic treatment than they have received.

The published work supports a clear but qualified position. Grid-forming control allows a small island system to hold frequency with few or no synchronous machines, and the main control families behave alike in normal operation because, at their core, they implement closely related laws. The decisive differences appear at the limits: when current saturates, when the energy source behind the converter is exhausted, and when fast converter controls meet slow diesel governors. An island utility assessing the technology is therefore better served by asking how a candidate unit behaves at those limits than by asking which control family it belongs to. Agreed specifications, validated models and published field records are the contributions that would most help it to do so.

To be declared by the author(s) before publication.

To be declared by the author(s) before publication.