If the baseline scenario was no timed commitments to the system operator, that's going to produce a dramatic claimed pricing benefit. But that's a bogus position, since no sane commercial wind farm operator operates without planning ahead. They could use public weather forecasts, most pay for commercially specific forecasts that concentrate on the wind forecast at turbine height (as well as forecasting power demand, since that's the other driver of wholesale price).
How does Deepmind do against the real baseline, of using a paid, commercial forecast that focuses on the details relevant to wind farm operators? I'm guessing something unimpressive like very low single digit percentages.