Predicting wind 36 hours out made it 20% more valuable to the grid
Google DeepMind · 2019–ongoing
What they did
Wind power's core commercial problem isn't generating it — it's that grids pay more for electricity delivered on a promised schedule than for whatever shows up unpredictably. DeepMind trained a neural network on weather forecasts and historical turbine data across 700MW of Google's central-US wind capacity to predict output 36 hours ahead, then used those predictions to make optimal hourly delivery commitments a day in advance.
What happened
Google and DeepMind reported the system boosted the value of the wind energy by roughly 20%, compared to a baseline of no time-based delivery commitments — making wind schedulable, and therefore worth more, for essentially the first time. It's a self-reported figure from Google's own fleet, not an independently audited industry benchmark.
The so-what
Not every AI win needs a chatbot or an agent loop — this one is a forecasting model plus an optimization layer, applied to the specific economic mechanism (schedulability) that determines what a unit of clean energy is actually worth.