New AI Runs Big Batteries Smarter, Making Power Cheaper

Big batteries — the kind that back up factories, data centers, and solar farms — are only as good as the software deciding when to fill and drain them. Today's standard approach works in two steps: first predict tomorrow's electricity prices, weather, and demand, then plan accordingly. But great predictions don't guarantee great decisions. And AI trained by trial and error tends to look only at the last few hours, missing longer patterns like daily and seasonal cycles.

The researchers tried a shortcut. Instead of predicting first and deciding second, their system learns to decide directly from raw data — past prices, weather readings, battery temperature, and how much charge is left. They taught it by imitation: a slower, math-based planner that effectively knows the perfect answer in hindsight demonstrated good decisions, and the AI learned to copy those instincts, only much faster.

Across five industrial sites, the AI captured 62 to 77 percent of the theoretical best possible performance — the score you'd get with a crystal ball. It beat both the standard forecast-and-plan software and the trial-and-error AI competitors. Crucially, it aged the batteries at nearly the same rate as its expert teacher, meaning smarter decisions didn't come at the cost of a shorter battery lifespan.

The catch: how well it transfers to brand-new locations depends on the site. Policies trained on busy, high-activity sites adapted most reliably; quieter sites were hit or miss. That means each new installation may still need some local training. Still, the win is real — cheaper power, longer-lasting batteries, and one less piece of software to trust.