Unit Commitment and Dispatch
Day-ahead against real-time, security-constrained economic dispatch, and the two-settlement system.
Chapter 8 described the operator sorting offers from cheapest to most expensive and accepting them in order. That describes what the answer looks like. It leaves out the fact that a power station cannot be switched on the way a light can.
A large coal unit takes many hours to bring from cold to full load, and the fuel burned getting there is spent whether the unit ends up needed or not. Once running it usually has a minimum output below which it cannot operate stably, so it occupies space in the stack whether or not the system wants it. It has a minimum run time, because a plant shut down after two hours suffers thermal stress its designers did not intend, and a minimum down time before it can start again. Its output can only change at a certain rate.
None of that fits in a merit order. A merit order answers how much each machine should produce, given that it is running. Somebody has to decide, in advance and without knowing tomorrow's weather, which machines are running at all. Those are two different problems, and every organised market solves them at two different times.
Commit first, dispatch later
The first problem is unit commitment: a set of yes-or-no decisions about which units are synchronised and available for each hour of tomorrow. The second is economic dispatch: given the units that are running, how much each should produce right now. Commitment is lumpy and forward-looking. Dispatch is continuous and immediate.
Both carry the prefix security-constrained, which is where a great deal of the cost hides. The operator does not solve for the cheapest way to serve load. It solves for the cheapest way to serve load that would still hold together if any single major component failed without warning. That is the N-1 criterion, and it means the system is permanently running a more expensive configuration than it needs at this instant, in exchange for surviving the loss of the largest generator or a critical line.
So the two calculations are security-constrained unit commitment, run for tomorrow, and security-constrained economic dispatch, run continuously. In PJM, day-ahead prices are produced for each hour of the following day, and real-time prices are recalculated every five minutes against conditions as they actually are.
Table 9-1: The two markets and what each decides
| Day-ahead | Real-time | |
|---|---|---|
| Decides | Which units run, and a schedule for each hour | How much each running unit produces now |
| Cadence | Once, for 24 hours | Every five minutes |
| Based on | Forecast load, forecast wind and solar | Measured load and output |
| Nature | Financial. Nothing physical has happened yet | Physical. Electrons are moving |
| Volume | The large majority of energy clears here | Only the difference from the day-ahead position |
Two settlements, one of which surprises people
The last row does more work than it looks. Under a two-settlement system, a participant is settled first on its day-ahead position at the day-ahead price, and then only on the difference between that position and what it actually did, settled at the real-time price.
A generator that sold 100 megawatt-hours day-ahead and produced exactly 100 in real time is completely indifferent to the real-time price, however violent it becomes. It sold at the day-ahead price and delivered. Only the deviation touches real time. The same is true in reverse for a retailer that bought its forecast load a day early and forecast correctly.
Two consequences follow. The first is that real-time price spikes hurt whoever was short in real time, meaning whoever failed to arrange beforehand, rather than the market at large. Uri was so damaging in Texas partly for this reason: participants who were short into a multi-day event had to buy at the cap for days. The second is that the day-ahead market, despite being a forecast of a physical event that has not happened, is the market where most volume clears and most hedging is done, which is why Chapter 22 spends its time there rather than in real time.
Committed, then told to stand down
Commitment happens against a forecast, and the forecast is wrong. Load comes in lower than expected, or the wind blows harder than the model said, and a unit that was committed yesterday for a genuine expected need turns out to be unnecessary today.
The unit still started. It burned fuel getting to minimum load and it cannot shut down again inside its minimum run time without damage. So it sits at minimum output, producing power nobody particularly wants, in a market where the real-time price may have collapsed to nothing. From the outside this looks like an error. It is the predictable cost of having to decide a day early, and every system pays some version of it.
The money the price cannot carry
This is where a genuine crack in the design appears, and it comes back in several later chapters.
A price per megawatt-hour can compensate a generator for producing a megawatt-hour. It cannot compensate a generator for the act of starting, because starting is not measured in megawatt-hours. A unit might be instructed to run for six hours at a price that covers its fuel perfectly well and still lose money on the day, because the price never contemplated the several tens of thousands of dollars it cost to get going.
Economists call this non-convexity, and the practical name is more revealing. Because the market-clearing price cannot recover lumpy costs, operators pay make-whole payments, also called uplift: a side payment outside the price, calculated after the fact, that tops a unit up to its as-bid costs when following instructions left it short.
Uplift is money that moves through the market without appearing in any price. It is socialised across load, so everyone pays it, and it signals nothing to anyone about where or when the system was short. A market whose uplift is growing is telling you that a rising share of what it costs to run the system has stopped being visible in the number everyone watches.
The amounts are small against total energy cost and the concept matters more than the total. Anyone reading a power market from prices alone is reading an incomplete account, and the missing part grows as the fleet gets more flexible and starts more often. Chapter 12 argued that the energy price may not pay enough to keep a plant available. This chapter adds that the energy price cannot express some of what the system already spends.
Why any of this is a trading concern
The commitment decision explains market behaviour that makes no sense from the merit order alone.
It explains why a forecast error moves a price more than the same quantity of physical change does. A revision that flips a large unit from committed to not committed, or the reverse, moves a lumpy block of capacity rather than a marginal megawatt. It explains why day-ahead and real-time prices diverge persistently rather than randomly, since one is computed against expectations and the other against reality, and the gap between them is a tradeable quantity in its own right. And it explains why the practical skill in short-term power trading is forecasting the operator's decision rather than forecasting demand, because the operator is the one deciding which machines exist tomorrow.