Artificial intelligence is highly adept at spotting patterns and making predictions that are much too small and subtle for humans to pick out
But all the patterns that AI extracts are historical. They all assume that the events in the future will be caused by, and will act out, the same things that happened in the past.
The recent past remains statistically a good guide to the near future. Contingency plans deal with the rest. Using the former better saves money and makes the latter *less* likely.
We have seen this with computerised trading: that all they can do is find a past pattern of actions and try to fit that to what is happening now and will continue into the future. AIs have no ability to understand when the rules have changed, or when new and previously unseen conditions need to be applied.
The UKs electricity generation often runs very, very, close to its limits in the winter. Mainly due to cost-cutting: why spend money on maintaining plant and excess capacity when it won't be used?
To employ AI to shave further percentage points and thereby run even closer to the limits simply reduces the margin for the unexpected. And being unexpected, you can't blame an AI for not spotting those patterns in the past.
A dangerous game.
It's more likely about better scheduling/forecasting than cutting any reserve.
Cover for the largest expected single generator failure were increased when Sizewell (nuke) and then Longannet (coal) tripped in close succession in 2008. Maybe better modelling would have had the increased cover in place *before* then and 500,000 people would not have lost power.
PS. BTW, I worked with low-latency traders. I suspect it doesn't work quite how you imagine.
Here was one I wrote up at the weekend:
Guess what could compute a daily forecast ready to upload to those phones and laptops, just for example, as well as some real-time polling?
Some of it could be based on the data used here:
1 Mole = 25 Cagey Bees