Agent Coordination via Contextual Regression (AgentCONCUR) for Data Center Flexibility

TL;DR

AgentCONCUR addresses grid-aware coordination with networks of spatially distributed data centers by learning a contextual affine policy $\phi(x)=\beta_0+\beta_1 x$ that maps readily available grid features to data-center task shifts $\varphi$, while preserving feasibility through a cost- and feasibility-aware training objective derived from an underlying bilevel optimization. The approach shifts the heavy computation to planning time and enables real-time, low-latency coordination with public/trusted contextual data such as prices and demands. NYISO-based case studies show substantial dispatch-cost reductions (varying with NetDC penetration and latency tolerance) and demonstrate that AgentCONCUR closely tracks ideal coordination while using a richer set of features than base regression. The work highlights practical considerations, including feature selection, training time, privacy implications, and possibilities for market integration and differentially private training in future work.

Abstract

A network of spatially distributed data centers can provide operational flexibility to power systems by shifting computing tasks among electrically remote locations. However, harnessing this flexibility in real-time through the standard optimization techniques is challenged by the need for sensitive operational datasets and substantial computational resources. To alleviate the data and computational requirements, this paper introduces a coordination mechanism based on contextual regression. This mechanism, abbreviated as AgentCONCUR, associates cost-optimal task shifts with public and trusted contextual data (e.g., real-time prices) and uses regression on this data as a coordination policy. Notably, regression-based coordination does not learn the optimal coordination actions from a labeled dataset. Instead, it exploits the optimization structure of the coordination problem to ensure feasible and cost-effective actions. A NYISO-based study reveals large coordination gains and the optimal features for the successful regression-based coordination.

Figures (8)

• Figure 1: Interfaces and notation of the power system, network of data centers (NetDC), and communication network between data centers and users.
• Figure 2: 11-zone New York ISO system with 5 data centers. The arrows show active virtual links for peak-hour, real-time coordination under different coordination solutions for the 20% NetDC penetration level and 100% maximum latency loss. The change of NetDC electricity demand is given as the average across the test dataset.
• Figure 3: UC costs throughout the 24-hour planning horizon for three latency loss bounds. The thick lines depict the average cost and the shaded area is the confidence band. The colored area is the period of Power-NetDC coordination. The cost resulting from coordination is in green, and the non-coordinated solution is in gray. The area between the curves is the cost-saving potential.
• Figure 4: Average electricity consumption of five data centers from 4 pm to 8 pm. The horizontal lines depict the constant, baseline consumption without any coordination. The bars are hourly consumption under coordination.
• Figure 5: Average NYISO dispatch cost across the testing dataset under different coordination models for the varying NetDC penetration level and maximum allowable latency loss. The area between the dashed lines defines the cost-saving potential for regression-based coordination.

Theorems & Definitions (1)

• Remark