The problem. A grid has to be built for its highest hour of demand. Those few peak hours call on the most expensive, most polluting standby plants, and they drive prices and capacity costs. Batteries can shave the peaks, but only if they still have charge left when the real peak arrives, and nobody knows in advance which hour that will be.
The swarm. Each battery (100 MW, 4 hours of storage, 10 batteries making 1 GW in total) decides every hour whether to charge, wait or discharge. It sees the current demand, the weather, the time, its own charge and the swarm's average charge. The team is rewarded for keeping the grid close to a flat daily level.
Grid demand with and without the battery swarm
Hourly demand on the ERCOT Coast zone (Houston area), in megawatts. Pick a week. The highest-demand week opens first. Hover or tap for values.
Swarm's stored energy that week
Average charge across the 10 batteries (0% = empty, 100% = full).
The hour that mattered
The season's highest demand came on Sunday 11 October at 4 pm (18,556 MW). The simple rule had discharged its batteries from midday, as soon as demand rose above the day's average, so it was empty by 4 pm and couldn't touch the peak.
The swarm learned to hold a reserve. It waited until 2 pm, then released energy through the four highest hours, cutting the peak to 17,608 MW.
All approaches compared
Whole test period (2,208 hours).
| Approach | Peak cut | Swings reduced | Peak-to-average |
|---|---|---|---|
| Swarm | 949 MW | 9.9% | 1.496 |
| Proportional rulesends just enough batteries to close the gap | 0 MW | 9.1% | 1.577 |
| Threshold ruleall batteries discharge when demand is high | 0 MW | 5.0% | 1.577 |
| No batteries | 0 MW | 0% | 1.579 |
The rules are good at smoothing ordinary days (9.1% against 9.9%) but miss the one hour that sets the system's cost. That is exactly the judgement the swarm adds.
Scales without retraining
The same model, trained with 10 batteries, ran unchanged with the same 1 GW split into 100 and then 1,000 smaller batteries. It delivered the same 949 MW peak cut each time, and choosing all 1,000 actions took under a millisecond on one CPU core.
Data and assumptions
- Demand: ERCOT Coast weather-zone hourly load, 2020. Weather: NREL solar and weather data for the same area.
- Batteries: 10 × 100 MW / 400 MWh, 90% round-trip efficiency.
- Learned on January to September, tested on October to December, which was never seen while learning.