Week 03 — The Weather Does Not Fail Independently

Correlated output, and why independent-generator arithmetic overstates a fleet's reliability.

Week 03 · correlation · 10 August 2026

Treating renewable generators as independent sources overstates reliability.

Property

correlation

Stance

Advances the claim

How

The course trace has no independent version; its wind is one shared signal.

Studio activity

Rebuild week 2's six rejected low periods under independent draws and compare.

Add enough independent generators and the law of large numbers does your reliability work for you. Add enough generators into one weather system and it does nothing at all.

How this advances the thesis#

It supplies the mechanism behind the event. A drought is not many small unrelated shortfalls happening to coincide; it is one cause expressing itself across a whole fleet. That is why the event is the right unit of design, and why fleet-level averages mislead in the same way annual averages do.

Independence is not a conservative assumption here. It is the assumption that deletes the object of study.

In the session#

We work on the course’s own trace rather than a generic fleet, because the trace makes the mechanism inspectable: its wind term is a shared synoptic signal applied across the whole fleet, plus per-hour noise. There is no independent-generator version of it. That is a property of how the data is generated, and you can read it in the generator.

So you build the missing half yourself. Construct twenty generators as independent draws with the same marginal distribution, then take the six low periods that week 2’s definition dial rejects for duration under the canonical definition, and ask how many of them still appear once the shared signal is removed.

The count is the exercise; the explanation is the point. Report which of the six you think survive independence, which do not, and what property of the generator decides it. We are not looking for a matching number — we are looking for you to say what correlation is doing to the duration of a low period, as opposed to its depth.

Its relationship to the canonical event#

The event is one contiguous sixty-one-hour block rather than sixty-one scattered bad hours, and correlation is the reason it has that shape. Remove the shared signal and you do not get a shorter event — you get no event, because a definition built on consecutive hours has nothing to bite on. The canonical definition’s span parameter is therefore only meaningful in a correlated fleet: week 3 is what makes week 2’s most sensitive parameter mean anything at all.

What next week needs#

Correlation raises an obvious defence: spread out further. Whether that works depends on how large the weather is compared with the grid, which is week 4 — and week 4 takes this week’s remedy apart rather than extending it.