A Novel Hierarchical Co-Optimization Framework for Coordinated Task Scheduling and Power Dispatch in Computing Power Networks
Haoxiang Luo, Kun Yang, Qi Huang, Schahram Dustdar · 2025
Read A Novel Hierarchical Co-Optimization Framework for Coordinated Task Scheduling and Power Dispatch in Computing Power Networks by Haoxiang Luo, Kun Yang, Qi Huang, Schahram Dustdar (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-04015.txt.
Introduction
Haoxiang Luo and four co-authors present a framework that schedules computing jobs and electric power together, so data centers run greener. Their paper came out in 2025. Computing power networks rent out computation on demand, but they burn immense energy and increasingly share grids with fickle renewable sources.
The framework works in two stages. A day-ahead stage plans generator commitments under uncertainty. A real-time stage then dispatches power while a learning agent shifts computing tasks toward hours and places where electricity is cheaper and cleaner. In simulations the system cut carbon emissions and running costs by clear double-digit margins. It also slashed wasted renewable energy while keeping nearly all tasks successful and fast. The authors pitch the coupling as a template for cross-domain optimization.
Catalog record
| CREATOR | Haoxiang Luo, Kun Yang, Qi Huang, Schahram Dustdar |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 46 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2508-04015
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
A Novel Hierarchical Co-Optimization Framework for Coordinated Task Scheduling and Power Dispatch in Computing Power Networks Haoxiang Luo, Kun Yang, Qi Huang, Schahram Dustdar Abstract The proliferation of large-scale artificial intelligence and data-intensive applications has spurred the development of Computing Power Networks (CPNs), which promise to deliver ubiquitous and on-demand computational resources. However, the immense energy consumption of these networks poses a significant sustainability challenge. Simultaneously, power grids are grappling with the instability introduced by the high penetration of intermittent renewable energy sources (RES). This paper addresses these dual challenges through a novel Two-Stage Co-Optimization (TSCO) framework that synergistically manages power system dispatch and CPN task scheduling to achieve low-carbon operations. The framework decomposes the complex, large-scale problem into a day-ahead stochastic unit commitment (SUC) 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 (DRL) agent. This agent makes intelligent, carbon-aware decisions by responding to dynamic grid conditions, including real-time electricity prices and marginal carbon intensity. Through extensive simulations on an IEEE 30-bus system integrated with a CPN, the TSCO framework is shown to significantly outperform baseline approaches. Results demonstrate that the proposed framework reduces total carbon emissions and operational costs, while simultaneously decreasing RES curtailment by more than 60 % 60\% and maintaining stringent Quality of Service (QoS) for computational tasks. I Introduction I-A Background The digital transformation of the global economy is fueling an unprecedented demand for computational power. The rise of artificial intelligence (AI), big data analytics, and the Internet of Things (IoT) has pushed traditional centralized cloud computing architectures to their limits [ 1 ] . In response, a new paradigm known as the Computing Power Network (CPN) has emerged [ 2 ] . A CPN aims to interconnect vast, geographically distributed, and heterogeneous computing resources, spanning from large-scale data centers to edge nodes, into a unified, programmable fabric [ 3 ] . The core vision of CPN is to break down resource silos and enable the flexible, on-demand scheduling and allocation of computing, storage, and network resources, thereby improving efficiency and enabling novel, low-latency applications. Concurrently, the global energy sector is undergoing a profound transition towards sustainability, marked by massive investments in Renewable Energy Sources (RES) such as wind and solar power [ 4 ] . This shift presents a powerful, yet challenging, symbiotic relationship with the burgeoning CPN paradigm. On one hand, CPNs are voracious consumers of electricity; data centers alone are projected to account for a significant portion of global electricity load growth, with AI workloads being a primary driver. Globally, data centers already account for 1-2\% of total electricity consumption, a figure comparable to the aviation industry. Driven by the explosive growth of AI, this demand is projected to double by 2030 [ 5 ] . In the United States, data center electricity usage is forecast to climb from 4.4\% of the national total in 2023 to as high as 12\% by 2028 1 1 1 https://www.energy.gov/articles/doe-releases-new-report-evaluating-increase-electricity-demand-data-centers , with AI workloads shifting from a minor component to a primary driver of this expansion. The availability of clean, renewable energy offers a direct path to decarbonize this massive computational infrastructure [ 6 ] . On the other hand, the very nature of CPNs, with their inherent flexibility in workload scheduling, presents a unique opportunity to support and stabilize a grid increasingly reliant on intermittent RES [ 7 ] . I-B Research Motivation The central challenge lies in a fundamental misalignment between the operational dynamics of CPNs and RES-dominated power grids. CPNs are designed to host a wide array of computational tasks, from latency-sensitive services like cloud virtual reality to compute-intensive, batch-processing jobs like large model training [ 8 ] , [ 9 ] . These tasks demand a highly reliable and stable power supply to ensure Quality of Service (QoS). However, RES are inherently variable, non-dispatchable, and dependent on weather conditions. This intermittency introduces significant volatility into the power grid, leading to challenges in maintaining frequency and voltage stability, and often results in the wasteful curtailment of clean energy when generation exceeds demand [ 10 ] , [ 11 ] . Simply connecting a CPN to a grid with high RES penetration without intelligent coordination creates a direct conflict. During periods of low RES output, the CPN’s demand would force the grid to rely on expensive and carbon-intensive fossil-fuel peaker plants to maintain balance. Conversely, during periods of high RES output, the grid may be forced to curtail wind or solar generation to prevent overload, even as CPNs continue to draw power from a mixed-carbon source. This decoupled operation leads to a suboptimal outcome: either unreliable computation or high operational costs and carbon emissions. The problem is not merely about sourcing green energy, but about managing the flexibility of CPN workloads as a grid-stabilizing asset. The inherent ability to shift CPN tasks in time (delaying non-critical jobs) and space (migrating workloads to regions with abundant RES) constitutes a powerful form of Demand Response (DR) [ 12 ] . This reframes the problem from ” how to power the CPN cleanly ” to ” how to leverage the CPN’s flexibility to enable a cleaner, more stable grid ”. This symbiotic relationship is the core motivation for the co-optimization framework proposed in this paper. I-C Contributions of the Paper This paper proposes a holistic, hierarchical co-optimization framework that treats the power system and the CPN as a single, integrated entity. By breaking down the silos between power system operators and CPN schedulers, the framework unlocks significant economic and environmental benefits. It coordinates power generation dispatch with computational task scheduling across multiple timescales to align CPN energy demand with the availability of low-cost, low-carbon renewable energy. As far as we know, this is one of the first efforts to jointly optimize the power grid and the CPN. The main contributions of this work are summarized as follows: • A Comprehensive Integrated System Model: A detailed model is developed that captures the intricate interplay between a heterogeneous CPN and a modern power grid. The model incorporates conventional thermal generators, stochastic RES, and battery energy storage systems (BESS), while explicitly modeling the spatio-temporal dynamics of electricity prices and carbon intensity. • A Novel Two-Stage Co-Optimization (TSCO) Framework: A hierarchical framework is designed to decompose the computationally intractable joint optimization problem into two manageable stages. A day-ahead planning stage addresses long-term unit commitment and resource reservation, while a real-time operational stage handles dynamic economic dispatch and adaptive task scheduling. • Scalable Optimization with Benders Decomposition: To address the large-scale, mixed-integer nature of the day-ahead Stochastic Unit Commitment (SUC) problem, Benders decomposition is employed. This technique effectively decouples the integer commitment decisions from the continuous dispatch variables, ensuring the problem remains computationally tractable even for large systems and numerous uncertainty scenarios. • Adaptive Real-Time Scheduling with Deep Reinforcement Learning (DRL): A DRL agent is developed for the real-time CPN task scheduling subproblem. This model-free approach enables fast, adaptive, and carbon-aware scheduling decisions in response to the highly dynamic and complex state of the joint CPN-grid system, complementing the model-based optimization