First-hand information for everyone
Modern large data centers running AI training create rapid, repeated swings in power use. Those swings can push voltages in the local electric distribution network outside safe ranges. This paper proposes a simple, local control method that takes advantage of a key fact about training: the total load tends to alternate between two dominant levels — a high-power compute phase and a lower-power communication phase. By aligning voltage targets with those two levels, the method reduces voltage swings while cutting the amount of corrective action needed.
The authors start from how power engineers normally control voltage: inverter devices adjust reactive power (a form of power that helps shape voltage) based on a fixed voltage reference. That standard approach, often called droop control, reacts after a voltage has moved and is tuned for slowly changing loads. The paper models data-center power as switching between two modes and shows that if the controller instead switches its voltage reference between two matching levels, part of the voltage change caused by a load step is absorbed by the reference itself. In the paper’s linearized network model, choosing the reference appropriately can roughly halve the worst post-transition voltage deviation compared with a fixed reference.
Importantly, the proposed controller is decentralized: each inverter updates its reactive power using only its local voltage measurement and a time-varying local reference. The authors derive conditions under which voltages converge and describe how to pick the two reference levels so they align with the data center’s two operating modes. They also develop a local switching rule so the controller changes reference based on measured voltage, not on direct coordination with the data center’s job scheduler. Simulation studies reported in the paper show substantial reductions in both voltage deviations and in the total reactive power changes (control effort) compared with conventional fixed-reference droop control. The approach is compatible with internal data-center measures such as battery smoothing.
This work matters because hyperscale AI workloads are becoming a new kind of fast, concentrated power demand on distribution networks. Frequent voltage corrections can wear equipment and raise operational cost. A control law that uses the predictable two-level pattern of training to avoid frequent large corrective actions can keep voltages within limits more efficiently, and do so without heavy coordination between the grid operator and the data center.
There are important limits to keep in mind. The method depends on the two-mode, synchronized behavior of modern training jobs; it is designed for loads that switch collectively between a compute phase and a communication phase. The analysis uses a linearized power-flow model and assumes other loads on the feeder are roughly constant; real networks and load mixes may be more complex. The results reported in this excerpt come from simulation studies rather than field trials. The paper also notes that the placement of the voltage references in the network matters and they characterize when placement is effective, so benefits are not uniform in every network.
Overall, the paper offers a practical, locally implementable tweak to inverter controls that exploits the structured timing of AI training. The authors provide both theoretical convergence conditions and simulation evidence that switching the voltage reference between two levels can reduce worst-case voltage spikes and the reactive control effort needed to manage them.