A Novel Hierarchical Co-Optimization Framework for Coordinated Task Scheduling and Power Dispatch in Computing Power Networks

The proliferation of large-scale AI and data-intensive applications has driven the development of Computing Power Networks (CPN). It is a key paradigm for delivering ubiquitous, on-demand computational services with high efficiency. However, CPNs face dual challenges in service computing. Immense energy consumption threatens sustainable operations. And the integration with power grids also features high penetration of intermittent Renewable Energy Sources (RES), complicating task scheduling while ensuring Quality of Service (QoS). To address these issues, this paper proposes a novel Two-Stage Co-Optimization (TSCO) framework. It synergistically coordinates CPN task scheduling and power system dispatch, aiming to optimize service performance while achieving low-carbon operations. The framework decomposes the complex, large-scale problem into a day-ahead stochastic unit commitment stage and a real-time operational stage. The former is solved using Benders decomposition for computational tractability, while in the latter, economic dispatch of generation assets is coupled with an adaptive CPN task scheduling managed by a deep reinforcement learning agent. It makes carbon-aware decisions by responding to dynamic grid conditions, including real-time electricity prices and marginal carbon intensity. Extensive simulations demonstrate that the TSCO outperforms baseline approaches significantly. It reduces carbon emissions by 16.2% and operational costs by 12.7%, while decreasing RES curtailment by over $60%$, maintaining a task success rate of 98.5%, and minimizing average task tardiness to 12.3s. This work advances cross-domain service optimization in CPNs.

💡 Research Summary

The paper tackles the emerging challenge of jointly managing large‑scale computing workloads and power system operations in the era of AI‑driven demand and high renewable energy (RES) penetration. Recognizing that Computing Power Networks (CPNs) are no longer passive loads but flexible, distributed computing fabrics tightly coupled with the electric grid, the authors propose a Two‑Stage Co‑Optimization (TSCO) framework that integrates day‑ahead stochastic unit commitment (SUC) with real‑time carbon‑aware task scheduling.

In the day‑ahead stage, the joint optimization problem is decomposed using Benders decomposition. The master problem decides binary on/off statuses of conventional generators, while sub‑problems solve continuous dispatch for each RES scenario, thereby capturing the stochastic nature of wind/solar output and its impact on electricity prices and marginal carbon intensity. This decomposition makes the large mixed‑integer stochastic problem tractable even for a 118‑bus test system with dozens of uncertainty scenarios.

The real‑time stage addresses the fast, highly nonlinear scheduling decisions required by CPN workloads. Each job is modeled as a Directed Acyclic Graph (DAG) with hardware‑type constraints (CPU, GPU, TPU) and data‑transfer requirements. A deep reinforcement learning (DRL) agent, built on an actor‑critic architecture, observes the current grid state (price, carbon intensity, available RES, storage levels) and decides where, when, and on which hardware each sub‑task should be executed. The reward function simultaneously penalizes (i) task tardiness and deadline violations, (ii) electricity cost, and (iii) carbon emissions, while rewarding the utilization of low‑carbon energy. To guarantee feasibility with respect to the physical power system, the Benders‑derived Lagrange multipliers are incorporated into the DRL reward, ensuring that the learned policy respects generation limits, network constraints, and storage dynamics.

Extensive simulations combine the IEEE 118‑bus system with 20 geographically dispersed CPN nodes, each equipped with heterogeneous compute resources. Workloads reflect realistic AI training and inference pipelines with diverse arrival times and deadlines. The TSCO approach is benchmarked against three baselines: (1) a conventional power‑only optimization with static CPN load, (2) a carbon‑aware CPN scheduler that ignores grid feedback, and (3) prior intra‑domain DRL schedulers from the literature. Results show that TSCO reduces total CO₂ emissions by 16.2 %, cuts operational costs by 12.7 %, and lowers RES curtailment by more than 60 % compared with the baselines. Importantly, it maintains a task success rate of 98.5 % and an average task tardiness of only 12.3 seconds, demonstrating that environmental benefits do not come at the expense of Quality‑of‑Service.

The contributions are threefold: (1) a comprehensive integrated model that captures both the physics of the power grid (including stochastic RES, battery storage, and network constraints) and the detailed resource‑constrained execution of CPN workloads; (2) a scalable hierarchical solution that leverages Benders decomposition for the slow, physics‑heavy planning horizon and model‑free DRL for the fast, stochastic scheduling horizon; (3) a demonstration that treating CPNs as flexible demand‑response assets can substantially improve grid sustainability while meeting stringent computing QoS requirements. The authors suggest future extensions toward multi‑regional market coordination, inclusion of electric‑vehicle fleets as joint compute‑storage resources, and integration of carbon‑pricing mechanisms, positioning TSCO as a foundational step toward truly carbon‑neutral, compute‑aware power systems.