Table of Contents

Designing a study

NemSim's published work so far asks one question: what does adding a large new load to the grid do to storage requirements, reliability and system cost. That is a narrow use of a general instrument.

The model is deterministic, every assumption is either a documented constant or a scenario input, and a run takes minutes rather than hours. Those three properties together mean the natural unit of work is not a run but a study: a set of runs that differ in one controlled way, published together with their provenance.

This section is about designing those.

The determinism contract

Everything here rests on one guarantee: the same inputs at the same commit produce identical results. There is no random number generator in the model. Rerunning a scenario reproduces every modelled value exactly; in the dispatch artifacts, the only field that moves is the runId stamped on each one, which identifies the run rather than describing it. Imported input artifacts and the sweep index also carry timestamps and durations, which Outputs and provenance sets out.

That is why a sweep is meaningful. When you vary one input across twenty points and the storage requirement moves, the input moved it. There is no run-to-run noise for the effect to hide in, and nothing to average.

It also means a result is checkable. A sweep records the git commit SHA, whether the working tree was dirty when it ran, and the SHA-256 of every input file it consumed. Anyone with the repository can reproduce your numbers exactly, or show that they cannot.

The corollary is that the model reports a realisation, not an expectation. It has no distribution to sample and gives no confidence intervals. See Limitations §1.

Comparative, not absolute

The single most useful habit when working with NemSim: read differences, not levels.

The model's biases are systematic. Transmission route length runs long. Greedy storage dispatch is wrong in a known direction. A typical meteorological year understates tail events. Every one of those sits on both runs when you compare two, and largely cancels.

So a claim of the form "storage requirement roughly triples between +2 GW and +6 GW of new load" is far better supported than "the system needs 14 GWh of storage". Frame findings the first way.

What is worth varying

Anything in a scenario config can be swept. These are the axes that repay the effort.

Demand

The published study. Add load and watch where the system stops absorbing it.

Vary dataCentreNameplateMw per region. Despite its name it is just a flat, always-on load increase, so use it for electrification, industrial load, or anything without a distinct shape. Watch for the point at which the sizing outcome switches from notRequired to resized, and later to batteryCapacityLimitReached or energyLimited. Those transitions are the interesting part; the smooth stretches between them are not.

Generation mix

Vary nameplateCapacityMw across fleets: more wind against less solar, retiring coal, adding gas.

The revealing outputs are curtailment and storage, not cost. A mix that looks cheap on capacity alone often turns out to need storage that the capacity comparison never showed, because solar and wind fail at different times of day and year.

Note that removing a fleet entirely and setting its capacity to zero are not the same thing: a zero-capacity storage plan is how you tell the sizing loop "build whatever this needs, and price it at these economics".

Economics

Vary costBasis.realDiscountRate, capital costs, fuel prices, or technical lives.

Discount rate first, always. It applies to every capital cost in the system, and its effect is asymmetric: raising it favours low-capital, high-fuel plant, and lowering it favours renewables and storage. A conclusion that flips between 5% and 9% was never a conclusion about technology.

Fuel price and heat rate are different in kind from capital cost: together they set short-run marginal cost, so changing them changes merit order and therefore what is actually generated, not merely what the same dispatch costs.

Reliability standard

Vary storageSizing.targetUsePercentage to price the standard itself.

Expect a sharply non-linear curve. The last increment of reliability is bought against the rarest events, so it is bought with the most storage. This is one of the more policy-relevant sweeps the model supports, and one of the least often run.

Transmission

Vary interconnector capacityMw, or add and remove links entirely.

Interconnection and storage are substitutes: both move energy from where and when it is abundant to where and when it is not. Sweeping one while the sizing loop solves for the other shows the trade-off directly. Read Limitations §6 before drawing a cost conclusion from it.

Designing a good sweep

One axis. If two things change between points you cannot attribute the difference. That is what a sweep is for.

Points where the interesting thing happens. A linear ladder wastes runs on the flat stretches. Run a coarse sweep first, find the region where behaviour changes, then sweep finely through it.

Include a baseline. A zero-change point makes every other point readable as a delta and catches the case where your overrides did nothing at all.

Push until it breaks. Extend the axis past the point where sizing stops succeeding. The outcome codes energyLimited, storageNoLongerImprovesReliability and batteryCapacityLimitReached are findings rather than failures, and the value at which one first appears is often the headline.

Make axisValue honest. It is a display value only; nothing in the model reads it. A sweep whose axisValue disagrees with its overrides runs happily and produces a chart that lies. See Sweeps.

Reading the result

A sweep publishes 19 scalars per point per region: levelised costs, energy served, renewable shares, storage capacity, unserved energy, curtailment and net imports. They are listed in Outputs and provenance.

Look for three things:

  1. Where the curve bends. Smooth response means the system is absorbing the change with what it already has. A bend means something became binding.
  2. Which component moved. Total levelised cost rising tells you little; generation cost rising while storage cost is flat tells you a great deal.
  3. What the sizing outcome says. A point that met its target with no new storage and a point that hit the capacity limit are qualitatively different results, even if their costs look similar.

Sensitivity analysis works a real example through.

Next

Edit this page