| 引用本文: | 周 雅,王 乾,方如举.配电网中基于混合DRL的任务卸载与多资源协同调度优化方法[J].电力系统保护与控制,2026,54(04):165-174.[点击复制] |
| ZHOU Ya,WANG Qian,FANG Ruju.Hybrid DRL-based task offloading and multi-resource coordinated scheduling optimization method in distribution networks[J].Power System Protection and Control,2026,54(04):165-174[点击复制] |
|
| 摘要: |
| (1.许昌学院，河南 许昌 461000；2.华北水利水电大学，河南 郑州 450045)
摘要：针对配电网在数字化、分布式和智能化演进过程中面临的“计算-通信-能源”多资源协同调度与任务卸载导致的时延-能耗联合最优化问题，构建了涵盖本地终端、边缘服务器与云端的数据驱动三层协同计算模型。该模型以加权时延-能耗-公平指标函数为优化目标，综合刻画无线信道条件、传输速率和CPU频率等关键因素，从而量化多资源协同对系统性能的影响。为应对离散卸载决策与连续带宽/计算/能量分配构成的混合动作空间挑战，提出混合深度强化学习(hybrid deep reinforcement learning, HDRL)框架。上层采用双重深度Q网络(double deep Q-network, DDQN)进行卸载动作选择，下层利用深度确定性策略梯度(deep deterministic policy gradient, DDPG)实现连续资源调度，并设计改进优先级经验回放机制(improved prioritized experience replay, IPER)提高样本利用率与收敛速度。仿真结果表明，与纯本地计算、纯边缘计算、随机卸载、遗传算法(genetic algorithms, GA)和不含IPER的DDQN+DDPG方法相比，所提HDRL算法在多场景下显著降低了系统平均时延与总能耗，同时，能在用户规模扩大时依旧能维持高公平性，表现出最佳的扩展鲁棒性，提升了任务完成率与算法稳健性，为配电网多资源协同优化提供了可行、高效的解决方案。 |
| 关键词: 边缘计算 任务卸载 资源分配 配电网 深度强化学习 |
| DOI：10.19783/j.cnki.pspc.250406 |
| 投稿时间：2025-04-25修订日期：2025-08-01 |
| 基金项目:国家自然科学基金项目资助(62103349)；河南省科技攻关项目资助(232102210104)；河南省研究生联合培养基地项目资助(YJS2024JD38) |
|
| Hybrid DRL-based task offloading and multi-resource coordinated scheduling optimization method in distribution networks |
| ZHOU Ya1,2,WANG Qian2,FANG Ruju1 |
| (1. Xuchang University, Xuchang 461000, China; 2. North China University of Water Resources and
Electric Power, Zhengzhou 450045, China) |
| Abstract: |
| To address the joint latency-energy optimization problem arising from task offloading and coordinated scheduling of “computation-communication-energy” multiple resources during the digitalization, decentralization, and intelligent evolution of distribution networks, a data-driven three-layer collaborative computing model encompassing local terminals, edge servers, and the cloud is developed. With a weighted delay-energy-fairness objective function, the model comprehensively characterizes key factors such as wireless channel conditions, transmission rates, and CPU frequencies, thereby quantifying the impact of multi-resource coordination on system performance. To tackle the challenge of a hybrid action space composed of discrete offloading decisions and continuous bandwidth, computation, and energy allocation, a hybrid deep reinforcement learning (HDRL) framework is proposed. In this framework, a double deep Q-network (DDQN) is employed at the upper layer to select offloading actions, while a deep deterministic policy gradient (DDPG) algorithm is used at the lower layer for continuous resource scheduling. An improved prioritized experience replay (IPER) mechanism is further designed to enhance sample utilization efficiency and convergence speed. Simulation results demonstrate that, compared with pure local computing, pure edge computing, random offloading, genetic algorithms (GA), and the DDQN+DDPG method without IPER, the proposed HDRL approach significantly reduces average system delay and total energy consumption across multiple scenarios. Moreover, it maintains high fairness as the number of users increases, exhibiting superior scalability robustness, improved task completion rates, and enhanced algorithm stability. The proposed method thus provides a feasible and efficient solution for multi-resource coordinated optimization in distribution networks. |
| Key words: edge computing task offloading resource allocation distribution network deep reinforcement learning |
