AI data centers get decision-focused battery scheduling: mean minute-to-minute grid-import swings fall 71.70% and total cost 18.01%
Synopsis
Power swings at AI data centers make grid-import smoothing hard, and with limited battery power and energy, forecast errors of similar size can lead to very different scheduling costs. BS-DFL writes the battery scheduling objective into training: a constraint-aware neural optimization layer first learns schedules without optimal-solution labels, then is frozen so scheduling costs and ramping risk train the demand predictor, and deployment needs no optimization solver. Over a 10,000-minute simulation using loads derived from measured AI workloads, BS-DFL cut the mean absolute minute-to-minute grid-import change by 71.70% and total scheduling cost by 18.01% relative to an MSE-trained predictor with a Gurobi solver, at an average decision time of 18.77 ms.
Interpretation
Battery scheduling for AI data centers no longer optimizes forecast error alone; the demand predictor is trained on scheduling cost and grid-ramp risk. Conventional forecasting minimizes statistical losses such as mean squared error without explicitly considering how forecast errors change downstream battery scheduling decisions. On a 10,000-minute simulation trace using loads derived from measured AI workloads, BS-DFL reduced mean absolute minute-to-minute grid-import changes by 71.70% and total scheduling cost by 18.01% relative to an MSE-trained predictor coupled with a Gurobi solver.
A constraint-aware neural optimization layer produces the scheduling decisions directly, so neither training nor rolling deployment calls an optimization solver. Earlier decision-focused methods largely rely on differentiable convex optimization layers, which do not directly handle scheduling with discrete decisions, while other approaches still need optimization solvers during training or deployment. The layer coordinates advance reserve commitments and demand-responsive dispatch through scenario-aware encoding and sequential constraint-aware decoding, is trained on the scheduling objective without optimal-solution labels, and is then frozen while only the predictor is updated; in simulation it averaged 18.77 ms per decision with a 64.07 ms maximum.
A worse predictor can still yield better scheduling, which shows the value of training the predictor for the downstream decision. BS-DFL's forecast mean absolute error is 240.66 kW, above the 179.97 kW of the MSE-trained predictor. Relative to the MSE-trained predictor paired with the frozen neural optimization layer, BS-DFL reduced mean absolute grid-import changes by a further 14.09% and total cost by 1.17%, lowering the P95 change from 0.0971 kW to 0.0104 kW.
Perspective
The demonstrated scope is a simulation of 10,000 one-minute decisions with 60-minute rolling horizons, using load profiles from a 10 MW, 2,840-node AI facility at 80% target occupancy and a 1 MW/1 MWh battery with 10–90% state-of-charge limits. The authors propose future work on learned scheduling strategies, electricity-price uncertainty, and robustness across battery configurations and operating conditions.
The maximum grid-import change under BS-DFL is still 203.92 kW; the authors attribute the largest changes to the sharp demand decline in Fig. 2(c) and note that a 60-minute prediction horizon may not yet see the following demand trough and recovery. BS-DFL's forecast mean absolute error is higher than the MSE-trained predictor's, and its reserve cost is 2,181 CU above MSE plus Gurobi, offset by lower ramp penalties.