NEM Constraint Watch

7-day outlook

Expected binding hours per market day (rhythm model: climatology, outage calendar, ST PASA). Constraints that bind around the clock sit at 24.

Forecast scoreboard

Binds within 60 minutes, last 2 days. Judged per constraint and dispatch interval against realized binding.

Binds during a clock hour, by forecast lead time, last 7 days. ST PASA is scored only over the constraints it models.

Model report

Latest walk-forward backtest (expanding windows, no leakage; refreshed weekly).

1. Why constraints exist

Every 5 minutes, AEMO's dispatch engine (NEMDE) picks which generators run, cheapest first. Electricity travels over physical wires, and wires have limits. Carry too much and lines overheat, or the grid loses stability after a fault.

Only 500 MW of the cheap power fits down the line, so the rest must come from local plants. A constraint is how that limit is written down as a rule NEMDE must obey. There are more than 15,000 of them, and every one has the same shape.

2. Anatomy of a constraint equation

The left side contains only things NEMDE can adjust: generator outputs and interconnector flows. Each factor says how strongly that unit affects the limit; a factor of 0.8 means backing the generator off by 1 MW relieves the line by 0.8 MW. The right side is the limit itself, recalculated every 5 minutes from things NEMDE cannot control: equipment ratings, live SCADA measurements, demand, and a safety margin.

3. Headroom and binding

The gap between the two sides is the headroom. A constraint spends most of its life ignored and only matters when the headroom reaches zero. At that point NEMDE deviates from the cheap plan: generators on the wrong side of the limit are backed down even though they are cheap (constrained off) and others are brought up in their place (constrained on). In the published data, a binding constraint is one with a non-zero marginal value.

4. Marginal value, the shadow price

The marginal value answers a precise question: if this constraint's limit were relaxed by exactly 1 MW, how many dollars would the next 5 minutes of dispatch save?

A constraint with slack headroom has a marginal value of zero, since relaxing a limit you are not touching saves nothing. That makes it the standard binding test. The sign depends on how the equation is written; the magnitude carries the meaning. Summed over time, marginal value multiplied by megawatts prices a piece of congestion, which is how candidate network upgrades get ranked.

One caution when reading dollar figures. In this project's data the median binding marginal value is about $0.38/MW and genuine congestion sits between roughly $0.10 and $1,000. A handful of constraints instead bind at their violation penalty, between $0.4M and $27M per MW. Those are hard operational rules, such as an interconnector forced to zero during an outage, and the number means "hold this at any cost" rather than a market price. The dashboard shows a penalty badge for these instead of a dollar figure.

5. How marginal value becomes regional price

A region's price is the cost of supplying the next 1 MW of demand there. Once a constraint binds, that next megawatt must be supplied without pushing the binding constraint past its limit.

1. Ask the price question in VIC: the next MW comes from the $30 plant, so VIC is $30.
2. Ask it in NSW: the $30 plant has spare capacity, but its path is full and the binding constraint forbids using it. The only legal source is the local $80 plant, so NSW is $80.
3. Ask the marginal value question: 1 MW more limit would let $30 power displace $80 power, saving $50.

The price gap and the marginal value are the same number seen from two sides. The marginal value is not a fee added to prices; it is the shadow cost of the bottleneck, and price separation is how that cost surfaces. For a simple interconnector limit, the importing region's price is roughly the exporting price plus the marginal value.

There is a subtler case when the binding constraint contains generators with factors. The next megawatt must then come from a blend of plants whose net effect keeps the left side exactly at the limit, and the resulting price can match no one's offer, exceed every offer in the region, or go negative. Per generator, a constraint moves its effective local price by the marginal value times its factor; summed over constraints, that is the local price adjustment AEMO publishes for each unit. The dashboard's expected-impact column estimates that exposure ahead of time.

6. Reading a constraint's name

This one reads: NSW thermal limit, in normal grid conditions, protecting line 39 against the trip of line 11. The cause symbols are > for thermal overload, ^ for voltage stability, : for transient stability and F_ for frequency reserves (FCAS). NIL means the grid is in its normal shape; when a line is out for maintenance, the outage token names it instead. The colored dots on the dashboard follow this classification.

7. How our models work

Hour ahead. For every enforceable constraint, every 5 minutes, a gradient-boosted classifier estimates the probability of binding within the next hour. Its evidence comes from four places: level and trend (headroom and how fast it is closing), rhythm (many constraints bind in daily patterns, so a constraint can carry hundreds of megawatts of headroom and still be likely to bind soon), context (time of day and AEMO's own 5-minute pre-dispatch view, including how many of its forward intervals bind and its projected closest approach to the limit), and outside drivers (regional demand, interconnector flows, renewables forecasts, temperature and the outage calendar). Raw scores are calibrated so the probabilities can be read at face value, and the alert threshold is fitted rather than guessed.

How hard will it bite. A second model predicts the likely marginal value of the upcoming bind using quantile regression on the same evidence. It replaced a simple historical-median estimate only after beating it by 28% on held-out error, and it is re-challenged at every retrain.

Week ahead. At that range binding is a rhythm problem. The weekly model is built on each constraint's hour-of-day history, with outage-calendar and AEMO ST PASA adjustments that switch on only when they improve validation accuracy. The 7-day outlook shows expected binding hours per day, which stays meaningful under any correlation between hours and does not saturate the way a chance-of-any-bind figure does.

Some things are deliberately left out: swamped equations (limits parked about 10,000 MW away, which is how AEMO disables a rule), informational rows that were published but never enforced, and auto-generated ramping constraints whose names change daily. Letting these in would corrupt both training and the watchlists.

8. How we grade ourselves

Every forecast this site makes is logged the moment it is made and scored once reality arrives. The Comparisons tab judges our model against AEMO's own forecasts (P5MIN, pre-dispatch and ST PASA) and a persistence baseline on identical terms, and a weekly walk-forward backtest is published as the model report. Some limits to keep in mind: unit-level aggregation approximates the correlation between constraints, a rhythm model cannot foresee a first-of-a-kind event that no outage calendar announced, and AEMO's pre-dispatch knows the actual forward limits inside about 40 hours, so beating it there is expected to be hard. The scoreboard shows where the crossover sits.

9. Data

Everything is built from public AEMO/NEMWeb data, refreshed every 5 minutes: dispatch actuals, P5MIN and pre-dispatch forecasts, ST PASA outlooks and constraint definitions from the monthly MMSDM archive, plus Bureau of Meteorology observations. Timestamps are NEM market time (no daylight saving). Predictions are a research model, not operational advice.