Decision-aware battery energy management in smart grids

Abstract

Energy management in smart grids increasingly depends on integrating battery energy storage, whose charge and discharge decisions shift demand, shave peaks, and capture price differences. Because every such decision consumes battery life, and warranties cap how many cycles may be used per year, the operator’s problem is deciding when battery use is worth its cost. This thesis proposes artificial-intelligence mechanisms for managing battery energy storage under that constraint, one at each of two time scales. At the planning scale, the thesis proposes a decision-aware approach to long-horizon price-spread forecasting for warranty-constrained cycle allocation, judging forecasts by the realized value of the allocation decisions they induce rather than by prediction error. Ontario motivates the problem: Market Renewal replaced the legacy province-wide hourly price with a day-ahead zonal price in May 2025, leaving post-renewal history too short to carry method claims, so a public New York Independent System Operator (NYISO) archive spanning 2000–2026 and eleven load zones provides the benchmark. The results show that the most accurate model is never the most valuable one in any of the eleven zones, and that selecting on validation allocation value rather than validation error yields higher held-out value in ten of them. This matters because accuracy-first selection, the default in electricity price forecasting, leaves battery value unrealized. At the operational scale, the thesis proposes a preference-conditioned multi-agent reinforcement learning controller that balances district peak reduction against battery cycling within one trained policy. Its actor uses peak-oriented and cycling-oriented specialist heads over a shared backbone, with an operator preference interpolating their outputs rather than entering the actor input, alongside structural control that keeps the executed actions feasible. In CityLearn with seventeen buildings over 30 days and three random seeds, the most peak-oriented setting cuts the 95th percentile of peak-hour district grid demand from 40.48kW to 33.41 ± 2.33 kW, a 17.4% reduction. The core contribution is one controller whose operating preference can change without retraining. Two supporting studies are kept separate from this result: an oracle-style forecast diagnostic examines discharge timing, and a limited language-interface demonstration maps operator instructions to the preference of the frozen policy. Neither study establishes a deployable forecast pipeline or general language-interface reliability.