The first half of this piece was about what "demand" actually means on the GB grid, and how little of it is measured. The short version: National Demand counts the output of everything NESO meters operationally — which means everything registered in the balancing mechanism, at any voltage. Everything below that line is not counted as generation at all; it just makes demand look smaller. About 15% of underlying demand is now estimated rather than metered.
That is a measurement problem, and measurement problems can be fixed by measuring things.
The question I get asked most about this — and lately it has been asked with a certain amount of heat — is some version of: how hard is it, really, to forecast the grid? Surely we can tell how much wind and solar will be running tomorrow?
Mostly we can, and the people who do it are good at it. But the question carries an assumption, and it is the one thing here that has stopped being true. Forecasting assumes you find out what happened. Every forecast is trained, tuned and scored against outturn: you learn what you got wrong, and that is how the next one gets better.
For the metered half of the fleet, that loop works. A transmission-connected wind farm's output arrives five days after the fact, per unit, from a meter. But for the categories growing fastest — the small sites, the rooftops, the panels behind a meter with a load sitting next to them — there is no outturn at all. Not late, not coarse. Absent. What gets published as yesterday's embedded solar is a model's estimate of yesterday, upscaled from a reference sample of systems that are metered.
It is a good estimate, built by people who know exactly what they are doing, and I said so in part one. But you can only score a forecast against it by comparing one estimate with another, and both lean on the same weather data. The error you measure that way is not the error you have.
So the honest answer to "how hard is it to forecast?" is that the forecast is the tractable part. The hard part is that for a growing slice of the system nobody ever establishes what happened, and it is difficult to systematically improve a prediction of a quantity that is itself inferred. You cannot converge on a number that nothing ever pins down.