Is VRE forecastable? (Part 1 in what might be a series)

… no, that’s not a trick question.

1)  In fact, it’s a million dollar (or actually many billion dollar) question – and one that’s central to the transition to a ‘VRE + Firming’ approach being adopted in the NEM (and, indeed, in many places without nuclear and/or deep hydro, and wanting to decarbonise).

2)  I’ve repeatedly seen on social media and elsewhere a number of stakeholders sprout claims that say either ‘no it isn’t’ or ‘yes it is’ … often without much substantiation included on either position (as if there were actually an absolute answer possible to the question, whereas ‘it depends …’ is probably a more).

At Global-Roam Pty Ltd  we we take the approach:

1)  that opinions aren’t worth much at all (and ‘old heuristics’ might end up increasingly outdated),

2)  but there’s a wealth of value of data in deep and reasoned analysis.

That’s why, for instance:

1)  we’ve invested a large amount of time and effort in compiling and recently releasing the GSD2024 (as the 7th annual update in the series), and

2)  why we invest so much time in building software tools like ez2view (complete with the ‘Forecast Convergence’ widget with it’s ability to allow one to ‘see that other dimension of time’).

But focusing on the question immediately at hand, we could answer:

1)  Certainly at a superbly superficial level, one might say that (for instance) ‘solar capability is eminently forecastable’, simply because the sun rises each morning and sets each evening

2)  But at a deeper level, we’ve certainly ‘put out money where our mouth is’ in making a sizeable investment in a talented start-up focused on wind forecasting (as we noted in this 25-year summary of our operations).

We would not have done this if we had thought that either:

(a)  it was a hopeless cause (i.e. to improve forecasts of VRE); or

(b)  that there was no real-world value to improving these forecasts.

… and already we’ve been pleased to see some results from our investment.

Our general level of busyness has meant that we’ve needed to call a temporary halt to the quarter analysis process feeding into GenInsights Quarterly Updates, but I’ll leave you with just two artefacts from the most recent report (for 2024 Q2 … all 304 pages of it) that speaks to the impressive challenges that remain to be solved in accurate forecasting of VRE.

About the forecastability of Large Solar

This first representation looks specifically at 2024 Q2, and a single chart that represents the compilation of ~36 different forecasts for each of 4,380 half hour periods across ~100 different Semi-Scheduled solar farm DUIDs through the quarter:

I don’t have time to go into the complexity that’s presented in just this image here:

1)  but two of the main takeaways that readers can understand here are that:

(a)   solar availability only really starts to become more ‘predictable’ in a mathematical sense is about 2 hours out from the point of dispatch.

(b)  and, a sizeable error remains even down to the point of dispatch.

2)  this obviously has big implications for the type of firming capacity required, the depth of storage, and how it will need to be ‘optimised*’ in this future environment.

* notwithstanding that ‘optimising’ this capacity might look different if viewed from the lens of a particular portfolio or participant and more altruistically from a ‘keeping the lights on services’ perspective … and also not going down considerations related to ‘the rise of the auto-bidder’.

About the forecastability of Wind

For all the Wind Farms across the NEM we’ve done the same analysis (updated each quarter for many of the quarterly updates across 9 editions).  In this extract we’ll present the results looking longitudinally with respect to the difference between the 5% and 95% percentile (i.e. another measure of ‘error’) as follows:

There’s two charts here, with the main takeaways including:

1)  With respect to the top chart, which shows absolute aggregate ‘error’, we can clearly see the ‘error’ continuing to grow as installed capacity of wind increases;

2)  With the bottom chart, which shows relative aggregate ‘error’ (i.e. relative to installed capacity) we can also see that:

(a)  The relative error did decline impressively through until ~2019:

i.  By more than half!

ii.  In other words there was some benefit from increasing diversity

(b)  But that any rate of decline has slowed to be almost non-existent since perhaps ~2020

i.  does this mean no more diversity benefit in forecastability?

ii.  does it imply the ‘low hanging fruit’ are gone, and the challenge to improve further is much trickier?

(c)  Now a lower relative error might be fine (looking at things one way), if we assumed that ‘Underlying Demand’ might grow at the same rate of growth in installed Wind Capacity

(d)  … but (barring the hydrogen hype coming true or other scenario like that) this is not the case and so it’s the absolute error that presents a more useful view of the underlying challenge.  In the report we’ve extrapolated forward to 2025 using modelled rates of growth of Wind Capacity, and the ‘error’ out that far is quite confronting.

3)  And of course this does not start to take into account other considerations, like congestion-based curtailment and so on…

Please reference your analysis in making claims

So our own take on the challenge (informed by deep analysis, and evidenced by a sizeable investment) is that there are considerable improvements that still need to be made … so don’t believe anyone who promotes an opinion (especially without substantiation) that is at either extreme:

1)  Don’t believe someone who claims ‘VRE can’t be forecast’;

2)  But also don’t believe someone who claims ‘it’s easy to accurately forecast VRE’ and/or ‘… that it’s a solved problem’ and/or ‘nothing to be concerned about here …’  and so on…