Automatic Generation Control AGC

Where the number comes from

The system operator's control centre continuously computes area control error — the deviation of actual interchange from schedule, plus a bias term on frequency deviation — and that error is what AGC exists to drive to zero. The correction is split across enrolled resources and pushed out as a command.

The form of the command differs by market: some operators send an absolute megawatt setpoint for the resource, others send a normalised position within the regulation capacity the resource cleared, which the plant applies on top of the five-minute economic base point it already holds (SCED in ERCOT, RTD in CAISO).

Either way the plant is being told where to be, not how to behave, and the cadence is short: roughly every 2-6 seconds in most North American markets, with PJM around 2 seconds, CAISO and ERCOT around 4, NYISO around 6. The economic layer above it still moves on five-minute intervals; AGC is the fast correction riding on that slower schedule.

The three letters are North American vocabulary, and the function has other names elsewhere, which matters when a spec written for one market is handed to a plant in another. In Continental Europe the equivalent product is the automatic frequency restoration reserve, defined in the ENTSO-E system operation guideline as a reserve activated by an automatic control device (SOGL Art 3(2)(99)) — the TSO's load-frequency controller occupies the same position in the hierarchy as an AGC master, and the European standard product on PICASSO is fully delivered in 5 minutes rather than in seconds.

In Great Britain the fast paid frequency products are droop-shaped: they answer measured frequency rather than a dispatched signal, so a plant qualified there is not thereby qualified to follow a signal anywhere. In Australia, regulation FCAS is delivered through AEMO's own AGC while contingency FCAS is a local response. Read the market rules before assuming a signal-following capability is what is being bought.

Obligation sits in two different places and they are easy to merge by mistake. NERC's BAL family of standards places control-performance obligations on the balancing authority — the entity running AGC — not on the individual resource. What binds the plant is the ISO or RTO tariff and market rules it registered under, plus the executed interconnection agreement.

In the FERC-jurisdictional US markets, Order 841 is what opened participation models to storage in the first place, and Order 755 is why the regulation payment has a performance component rather than paying a sluggish unit and a fast one the same way for the same megawatt.

Commanded and remote versus autonomous and local

The clearest way to hold the distinction is to follow the path each response travels. An AGC response starts at telemetry arriving in a control centre, passes through the operator's computation, crosses a wide-area link to the plant's market gateway, is passed down to the plant controller, and is finally split across the PCS fleet — a chain of hops measured in fractions of a second to seconds, on top of a command that was only issued every few seconds anyway.

A droop response starts and ends inside the fence: the converter measures frequency and moves, and the proportional law that governs how far it moves belongs to the P-f droop entry, while the products built on that law belong to frequency response. Cut the communications link and the droop response is unaffected; the AGC response stops, which is exactly why the interconnection agreement has to state what the plant does on loss of the utility link rather than leaving it to firmware defaults.

They compose rather than compete, and the composition has a commercial edge to it. The AGC setpoint is the base the plant is holding; any mandatory primary frequency response is added on top of it, so the megawatts of headroom consumed by the obligation are not available to sell into a regulation product.

ERCOT, for instance, requires primary frequency response from generation resources and energy storage resources as a condition of connecting, with governor droop of 5% or less and a deadband no wider than ±0.017 Hz — uncompensated work that still eats headroom and throughput. A model that books the same megawatt as both regulation capacity and grid-code compliance fails its first serious review.

Telemetry, and the score attached to it

Signal following is a two-way arrangement: the operator cannot dispatch what it cannot see, so enrolment carries a telemetry obligation running at a cadence comparable to the command — commonly the same ~2-6 second range in North America.

The point list is market-specific, but the recurring items are net active power at the point of interconnection, resource status and availability, the regulation range currently available up and down, and for storage the state of charge or remaining energy in each direction, because a battery's ability to answer the next command is a function of where it sits. This is control-grade telemetry and it is not the settlement path: the quantity that gets invoiced still comes from the revenue meter and its accuracy class, which is that entry's subject.

What the telemetry feeds, besides dispatch, is a performance measurement. Under FERC Order 755 the FERC-jurisdictional US markets pay regulation on capability plus performance rather than capability alone; PJM, as one worked example, builds a score from accuracy, delay and precision terms and multiplies the payment by it, so tracking quality is revenue rather than a report-card entry. Two failure modes show up in that score before they show up anywhere else.

One is latency anywhere in the command path — a plant that is always answering the previous setpoint scores poorly however accurate its final value. The other is state of charge: the signal is only roughly energy-neutral, so a run of net-up or net-down movement drags the plant toward a rail, and a plant near either end can no longer answer in both directions. Operators manage that with SOC bands, market energy transactions to re-centre, and in some markets explicit rules about how storage restores its state of charge.

What the plant has to do with the setpoint

Four things, in order. Accept it: the signal terminates at the market-facing gateway and is passed inward through the control hierarchy, and which box owns that termination — RTU, plant controller, EMS — is a scope question worth settling in writing rather than at commissioning.

Track it: move to the new value before the next one arrives, referenced at the point of interconnection net of transformer, collector and auxiliary losses, which is the plant controller's job and the reason plant control cannot be open-loop; tracking tolerances of roughly 1-2% of the commanded value at steady state are common, though the binding figure is whatever the market rules and the witnessed test plan state.

Respect the envelopes: the interconnection limit, the PCS apparent-power rating, the SOC band and thermal headroom all outrank the command, and a setpoint that would breach one gets clipped — with the clipping visible in the performance score, which is the honest early warning that the plant is enrolled beyond what it can deliver. Report back, and fail to a pre-agreed state if the link drops.

On the paperwork side, AGC signal-following runs are normally part of the witnessed commissioning test plan and a prerequisite for market qualification, sitting alongside the P and Q capability tests.

The questions worth asking a vendor are narrow and answerable: what is the latency budget from signal receipt to measurable response at the POI, what happens to the plant when the utility link or the individual PCS links drop, which box arbitrates when the AGC command and an internal SOC-management action disagree, and how are setting and firmware changes governed after the interconnection model has been validated.

What signal-following duty costs

Start with the arithmetic, because it is the part people underestimate. A 2-second cadence is 1,800 setpoints an hour and 43,200 a day; a 4-second cadence is 900 an hour.

Each individual move is small: regulation assets are commonly sized at a 15-minute to 1-hour energy-to-power ratio, set by market rules rather than by chemistry, and one 2-second step at full commanded output moves 2/900 of the energy a 15-minute-duration plant holds at rated power — about 0.2%, or about 0.06% on a one-hour plant. So the duty is shallow and nearly continuous, the opposite shape from the one-deep-cycle-a-day profile most warranty tables were built around.

That shape is what has to be reconciled with the warranty, and the reconciliation is contractual before it is technical. Cell wear tracks energy throughput and the conditions it happens under; how a given profile ages the asset is the duty profile entry's subject, and the accounting conventions are energy throughput's. What belongs here is the question to put to the supplier: what does the warranty actually count — equivalent full cycles, MWh of throughput, or a cap on cycles per year — and does the duty profile it warrants against resemble signal-following at all?

Those three counting methods do not treat 43,200 small reversals the same way, and a warranty negotiated around a daily 4-hour cycle is not a statement about regulation duty until someone converts one to the other. Check as well whether a C-rate limit or a temperature condition in the warranty is compatible with the enrolled service.

The second cost line is efficiency, and it is a cost rather than a degradation footnote. Every MWh that jitters through the battery pays the round-trip toll, and the auxiliary load runs whether the plant is moving or idling. Regulation revenue is quoted gross; the energy bought back to hold state of charge, and the losses on the way through, are separate model lines. Neither of them usually breaks the business case — but a model that omits them flatters the margin on precisely the service that is easiest to over-enrol.