Table of Contents

Limitations

Read this before you quote a number from NEMSweep.

These are not disclaimers. Each one changes how a result should be read, and in most cases the direction of the error is known. They are stated ahead of the assumptions register, because a reader who takes only one page from this site should take this one.

Some limitations are the framework's (greedy storage dispatch, no unit commitment, a near-frontier sizing search). Others belong to the published example (a single typical weather year, operational demand rather than total demand). Both kinds are here, because both change how a published figure should be read.

The four that matter most

If you read nothing else on this page, read these four. Each links to the section that works it through.

Limitation What it does to a result
1. A deterministic run cannot honour an expectation-based standard Storage is understated. The standard is an expectation across many weather outcomes; one typical year is a realisation, and it excludes the tail events that drive reliability. Treat a sizing result as a lower bound.
2. Renewable share is measured on operational demand Not comparable to a rooftop-inclusive target. Operational demand nets out rooftop PV before the model sees it, and no rooftop fleet is dispatched. Two shares are published, on grid-scale and native bases; neither is "the" number.
3. This is the cost of building the system, not a power bill Not a price. The figures are annuitised build-and-run cost per MWh served. No distribution, retail, market, environmental scheme or tax costs are modelled.
4. Greedy storage dispatch is wrong in both directions Biased both ways. With no foresight, storage undersizes against a multi-day wind drought and oversizes against a system able to arbitrage. Which dominates depends on the scenario.

Four more follow, on sizing, transmission distance, interconnector losses, and what is not modelled at all.

1. A deterministic run cannot honour an expected-unserved-energy standard

The reliability standard NEMSweep sizes against, 0.002% of demand energy unserved, is an expectation across a distribution of weather and demand outcomes. NEMSweep runs a single weather year and produces a realised figure from it.

These are different quantities. Reliability is driven by tail events: the still, cold week; the wind drought that coincides with a heat wave. A typical meteorological year is assembled from representative months and, by construction, contains few of them.

Direction of the error: storage is understated. A system sized to meet 0.002% USE against one typical year will need more storage than that to meet 0.002% USE in expectation. Treat a sizing result as a lower bound on required capacity, not an estimate of it.

Doing this properly means running many weather years, or synthetic years drawn from a fitted distribution, and taking the expectation across them. NEMSweep does not do that today.

2. Renewable share is measured on operational demand, not a rooftop-inclusive basis

Renewable-generation targets are commonly quoted against total demand, including the part rooftop solar serves. The published example dispatches operational demand, which is demand as seen by the grid after rooftop PV has served part of it. Rooftop generation is netted out of the demand series before the model sees it, and no rooftop fleet is dispatched.

So a renewable share measured on what NEMSweep dispatches is not directly comparable to a rooftop-inclusive target. NEMSweep reports two shares, and neither is presented as "the" number:

Measure Numerator Denominator
Grid-scale renewable share Delivered Solar + Wind + Hydro Total delivered generation
Native renewable share Delivered Solar + Wind Base demand energy, excluding additive demand components

If you are comparing against a published target, check which basis that target uses. Comparing a grid-scale share against a target that counts rooftop will understate progress; the reverse will overstate it.

3. This is the cost of building the system, not a power bill

The levelised figures NEMSweep publishes are system build-and-run costs per MWh of energy served: annuitised capital plus one year of operating cost for the generation, storage and transmission assets in the scenario, divided by energy served (demand minus unserved energy).

A retail electricity bill is a different thing entirely. It also contains distribution network charges below the transmission level, retail operating costs and margin, market and settlement costs, environmental scheme costs, metering, and taxes. None of those are modelled.

Do not read a NEMSweep AUD/MWh figure as a price, a tariff, or a forecast of what anyone will pay. It is an engineering-economic cost of supply, and it is a modelled estimate rather than an audited one.

4. Greedy storage dispatch is wrong in both directions

NEMSweep's storage policy has no foresight. It sees the current interval only: the residual after generation has met demand, and each fleet's headroom. It cannot pre-charge in anticipation of tomorrow's shortfall because no forward view of residual demand is provided to it. It does not hold stored energy in reserve for forecast events.

That cuts two ways, and which way dominates depends on the scenario:

  • Against a multi-day wind drought, it undersizes. A policy that discharges whenever there is a deficit will spend its stored energy on the first day of a lull rather than rationing it across three. Real operators, seeing a forecast, would hold back. Storage sized against greedy dispatch is therefore sized against a worse operator than the real one.
  • Against a system able to arbitrage, it oversizes. A greedy policy charges only from surplus or from incremental thermal generation in the current interval. It never charges cheaply now to displace something expensive later. A system that could arbitrage would extract more value from the same megawatt-hours, and would need fewer of them.

The same absence of foresight is what makes the model deterministic and cheap to run, which is what makes sweeps practical. It is a deliberate trade, but it is a real limitation.

5. Sizing finds a near-frontier point, not an optimum

The storage sizing search grows Battery capacity until every region is inside the reliability target, then refines each changed region's power and energy coordinate by coordinate. While regions fail, it grows one at a time in descending total unserved MWh, with ordinal region-ID tie-breaking. It re-dispatches and re-ranks after each accepted change. A region's best candidate is accepted only if it reduces that region's own unserved energy by more than 0.001 MWh; blocked regions are skipped while other failing regions are tried, and reconsidered after an accepted change. Once all regions pass, refinement remains ordinal, power first and then energy. The result is a deterministic coordinate-wise near-frontier point.

It is not a global minimum, and it is emphatically not cost-optimal: nothing in the search prices the capacity it adds. The main-growth priority is intentional and can change successful sizing results compared with growing regions in ordinal order. A different search order could land on a different but equally compliant point. Two sizing results are comparable to each other because the procedure is deterministic; neither is "the answer".

A sizing run that does not meet the target reports the final blocker for each failing region. If no change is accepted, its aggregate outcome is batteryCapacityLimitReached when any failing region has no larger candidate, otherwise storageNoLongerImprovesReliability; reaching just one dimension's ceiling is not enough when a larger candidate remains on another dimension. Evidence lists all blockers. These outcomes, like a pass-budget stop, are bounded-search evidence, not a global infeasibility or optimality result. Only the energyLimited outcome, where total available generation energy across the system is below total demand energy, proves that no Battery could have met the target under that energy deficit.

The same applies to inter-regional transfer, which serves regions in priority order by successive max-flow solves. That guarantees a higher-priority region is never starved by a lower-priority one, and for the same reason is deliberately not a global optimum.

6. Reciprocal interconnectors are each costed at the corridor's full length

Every corridor that carries flow both ways is declared as two directed interconnectors (see Scenario configuration), and each is costed independently over the corridor's full route length, so the same kilometre of conductor is paid for twice. Charging each of the five physical corridors once at its larger directed rating gives 1,955,280 km·MW against the 3,461,890 km·MW actually charged across all ten directed links. This convention alone accounts for roughly 1.8× of reported transmission cost. It is kept deliberately: attributing a shared asset's cost to one direction rather than the other would need a convention of its own, and this one is at least uniform.

Route length itself is a declared field on each scenario interconnector (routeLengthKm), not derived from anything else. It used to be the great-circle distance between the endpoint regions' solar weather sites, which ran long against the real NEM corridors and meant swapping a region's solar weather file silently changed transmission cost. The committed scenario now carries approximate real corridor distances instead (Heywood ≈275 km, VNI ≈300 km, the QNI corridor ≈500 km, Basslink ≈370 km, EnergyConnect ≈900 km); a more precise per-route figure can replace any of them without a code change.

7. Interconnector losses are a flat unsourced placeholder

Every hop of an inter-regional transfer loses a flat 5%, so a two-hop route delivers 0.95². That figure is a placeholder. AEMO publishes marginal loss factors per interconnector and they are neither flat nor equal across links; a sourced value should replace this one.

Any result that depends materially on inter-regional transfer inherits that uncertainty.

8. What is not modelled at all

Stated so their absence is not mistaken for a zero:

  • No market. No bidding, no offers, no spot price, no settlement. Generation is dispatched in merit order by short-run marginal cost.
  • No unit commitment. No minimum stable generation, no start-up cost, no ramp rate, no minimum up or down time. A fleet can go from zero to full output in one hour.
  • No forced outages or maintenance. Every fleet is available at nameplate whenever its resource allows.
  • No intra-regional network. A region is a single node. Distribution and sub-transmission constraints do not exist.
  • No frequency control, inertia, or system strength. Reliability here is an energy measure only.
  • No capacity expansion over time. No multi-year build path and no retirement schedule. The cost engine prices exactly one year, and storage sizing solves for a single end state.
  • No demand response or price elasticity. Demand is exogenous and inelastic.

Reading a result honestly

The safest use of NEMSweep is comparative. Because the model is deterministic and every assumption is either a documented constant or a scenario input, the difference between two runs is trustworthy even where the absolute level is not: the same biases sit on both sides and largely cancel.

That is what makes sweeps the natural unit of work here, as Sensitivity analysis works through.

Next

Edit this page