Are Google Deepmind's weather forecasts any good for energy traders?

Nearly 5 years since a blog from Google's Deepmind announced that machine learning had boosted the value of its wind farms by 20%, they published an article last week in Science claiming that their GraphCast AI model "marks a turning point in weather forecasting".

Mistakenly described as "peer-reviewed" in Wired, the magazine adds to the almost universal acclaim of their claims of revolutionary capabilities.

Google Deepmind still refers to the 2019 blog, and apart from a mention in a summer 2023 piece entitled Using AI to Fight Climate Change, I have seen no other reference or update to the claims.

This mentions a collaboration with ENGIE to develop a software product using the model, and likewise, nothing but silence since the June 2022 press release.

Every big announcement is widely reported in the media, but are any traders actually using this algorithm to trade?

How does AI increase the value of wind energy?

The premise was that using a custom AI tool to better predict wind power output and another model to recommend commitments to the electricity markets based on the forecast would improve the financial return.

But the chart they shared of their wind forecast simply doesn't look that good:

Covering 3 days of forecasts, actual output was consistently higher then predicted, with Saturday afternoon generation double the prediction.

So if the wind forecast wasn't great, was the increase in value down to the model recommending the trading strategy?

The hypothesis is that selling day ahead would get a better price than leaving it to the imbalance price. The skew in the forecast error hints that their prediction is more likely the minimum output they should commit into the day ahead market, with any upside sold at the imbalance price.

In markets with dual imbalance pricing, there is typically a punitive spread between the buy and sell prices to encourage balance responsible parties to ensure their traded commitments match their expectations - a balanced position.

European markets are moving towards a single imbalance price that rewards generators and retailers for having an imbalance position that reduces the system imbalance. If the system is short, excess power will benefit from a high price. But if the system is long, excess power may face low or even negative prices.

It is better to be long and wrong, than short and caught

The risk of a negative imbalance position is much greater than that for a positive imbalance. This asymmetry reflects the impact of blackouts being much worse than having to curtail some generators. Prices spike much higher when there is insufficient power to meet demand than they fall when there is an excess.

The patterns that Google identified appear to be that it would be better to sell wind energy day ahead than at the imbalance price, but much worse to sell more than is actually generated. This is optimised by selling forward only the minimum generation that can be confidently predicted.

The 20% improvement in value was described as early results for an algorithm that would be further optimised. And their new weather forecasting models - Graphcast or the short-term MetNet-3 model - should also bring greater accuracy to the wind forecast. An update feels long overdue.