I can speak for the GB case. Low Frequency Demand Disconnection (LFDD) occurs automatically and in stages when the frequency drops until it stabilises. The substations or feeders that are tripped off are not currently determined by real-time metering - instead they are pre-allocated based on their typical demand. This means that the system operator does not really know how much demand will be disconnected at any given time. If it's sunny, you could easily trip off a lot of solar generation connected on the low voltage network, causing the frequency to drop further. It is far from optimal!
What's the comparison with conventional attention using a more aggressive (lower temperature) softmax? I can imagine that for the multi-needle retrieval test this may also give a performance boost, although at some cost other more creative tasks.
The vast majority of wind that is unharvested is curtailed due to constraints on the transmission network between Scotland (lots of wind) and England (lots of demand). There is a significant amount of pumped hydro (and battery storage) in Scotland to help in these instances but there are still enormous costs associated with curtailment (approx. £1 billion per year).
My first thought on seeing the PCA embeddings scatterplot was "I wonder what pdfs are at the centre of those two clusters?" The most typical pdfs on the internet.