A neural controller generated 3,650 day-ahead schedules for home batteries in 62 seconds on a central processor. A conventional optimisation benchmark took about 9.7 hours on the same machine. Every schedule produced by the filtered controller stayed within the operating limits, while its combined electricity and battery-wear cost was 3.33% higher. Johns Hopkins University and University of Nevada, Reno researchers reported the work in an unreviewed preprint on September 11.
A home battery has to choose when to charge, discharge or wait as solar output, household demand and prices change. The cheapest immediate action may also deepen a charge cycle and shorten battery life. Conventional optimisation can represent these choices accurately, but repeatedly solving the full day for many homes is slow.
The proposed system, MI-DPC, receives 24-hour forecasts and chooses a mode and power level for each 15-minute interval. During training, it learns both energy cost and wear estimated from the depth of each charge cycle. During use, a safety filter checks the planned state of charge and reduces power whenever the battery would cross a limit.
The researchers replayed one year of measured solar and demand data from ten homes in California, Colorado and Arizona. One model was trained across the whole fleet. Its schedules were compared on the same processor with a piecewise-linear optimiser, which served as a practical reference because finding the exact global optimum at this scale is infeasible.
Across the 3,650 battery-days, the reference schedules cost $16,000 and MI-DPC's schedules cost $16,533. The neural system was 564 times faster, and its safety filter left no constraint violations. A model trained on only one California home transferred less reliably, crossing charge limits on 457 days before filtering and producing a larger fleet-wide cost gap.
The study replayed historical data rather than controlling physical batteries. It used one battery chemistry and assumed that the daily forecasts were correct. A field trial with unseen homes and forecast errors must measure actual energy bills, capacity loss and safety-filter interventions before the speed gain can be weighed against the remaining cost difference.
Related coverage
- Grid Forecast AI Holds Error to 15.61 Amps in Simulated Data Attack. A grid-control study examines whether an AI forecast can remain usable when sensor data are manipulated.
Sources
- Eshagh Safarzadeh Ravajiri and colleagues, arXiv, September 11, 2026. Abstract, authorship, submission date and headline results.
- Full paper. Controller sequence, battery data, comparison methods, timing, feasibility and limitations.
- Lead image: Yan lance, “Household energy storage system”, released under CC0 1.0. The photograph is illustrative and does not show the experiment.
