Can Generic AI Predict Electricity Prices Better Than Custom Models?

Electricity prices change every hour, and if you own a big battery — or your utility does — you can make money by charging when power is cheap and selling when it's expensive. Predicting those prices accurately is worth a lot. For years, companies have used forecasting models trained on each specific market, like Germany's or Spain's. But a new breed of AI, called foundation models, wants to do the same job with little or no special training. Think of them as a Swiss Army knife instead of a custom-built tool.

Researchers from Wrocław University of Science and Technology tested nine foundation models against two specialized benchmarks in three European countries over four years. They found that one family of models, TabPFN, was consistently better at predicting the next day's electricity prices. It outperformed the specialized models in every market they tested. But here's where it gets interesting: being statistically better doesn't always mean being richer.

The team also simulated making money by buying and selling electricity with a battery. Under aggressive trading with unlimited bids, TabPFN made the most profit. But when they limited risk — like a cautious investor might — a specialized model actually earned more. That means the "best" model depends on how you plan to use it, not just its raw accuracy.

What's the takeaway for ordinary people? As power grids use more renewables, prices swing more wildly, and better predictions help lower costs for everyone. But don't expect one AI to rule them all just yet. The researchers advise matching the model to the real-world decision — whether that's charging your EV at the cheapest hour or helping a utility store solar power for the evening rush.