AI Runs Giant Underground Batteries 100,000x Faster Than Old Software

Engineers have taught an AI to operate an underground pumped-hydro plant. Here's the simple version of what that is: a giant battery made of water. When electricity is cheap and plentiful, you pump water up into a high reservoir. When everyone needs power, you let it rush back down through turbines. An underground version does this in deep shafts or retired mines, so it doesn't need a mountain and a valley. Grid operators must decide every day exactly when to pump and when to release — a planning puzzle with real money attached.

The puzzle is hard because it mixes two very different kinds of decisions. Some choices are continuous, like how fast to run a pump. Others are all-or-nothing, like "this unit is either on or off." On top of that, the machinery doesn't behave in neat straight lines — its performance curves are lumpy and complicated. Traditional solvers grind through all of that slowly. The researchers replaced the solver with a neural network (an AI trained on examples rather than hand-coded rules) that learns to spit out good schedules almost instantly.

Their results are striking. The AI's daily plan came within 1.6% of what the slow, trusted solver produces — close enough that most operators wouldn't notice the difference — while running about 100,000 times faster. That speed gap is the real story. A plan that once took hours can now be redone in the time it takes to blink, which matters because weather forecasts, demand and power prices shift constantly. Being able to re-plan in seconds means more cheap renewable energy gets used instead of wasted, and fewer dirty backup plants need to switch on.

There are honest caveats. This was a simulation study, published as a conference paper, not a plant running in the field. The 1.6% gap is small but not zero, and regulators will want proof before trusting an AI with critical infrastructure. The AI also needs a good model of the specific plant it's controlling. Still, the direction is clear: as grids fill up with solar and wind, the winners will be whoever can plan fastest.