模仿学习引导的深度强化学习用于云边协同光电数据流多维资源在线调度
DOI:
CSTR:
作者:
作者单位:

广东电网有限责任公司

作者简介:

通讯作者:

中图分类号:

基金项目:

2025年数字技术平台适应性改造(三区数据中心(省侧)能力升级)项目(037800HK24120111)


Imitation Learning-Guided Deep Reinforcement Learning for Online Multidimensional Resource Scheduling of Cloud–Edge Collaborative Optoelectronic Data Streams
Author:
Affiliation:

Fund Project:

  • 摘要
  • |
  • 图/表
  • |
  • 访问统计
  • |
  • 参考文献
  • |
  • 相似文献
  • |
  • 引证文献
  • |
  • 资源附件
  • |
  • 文章评论
    摘要:

    随着光纤传感、光电互感器等光电感知设备在电力边缘侧的规模化部署,海量高带宽光电感知光电数据流对云边协同架构的实时传输与处理构成了严峻挑战。针对电力云边协同光电数据流处理中资源异构、任务依赖复杂、多目标冲突的在线调度难题,研究将调度过程建模为马尔可夫决策过程,并提出模仿学习引导的深度确定性策略梯度(IMITA-DDPG)调度算法。该算法分层融合模仿学习与深度确定性策略梯度,依托历史最优调度日志训练专家策略,以引导智能体在光电光电数据流高动态场景下实现高效探索与稳定优化;同时,构建了多目标协同的目标函数与四层约束体系,实现任务卸载、内存分配与带宽预留的一体化协同决策。结果表明,该算法在Sphere和Rastrigin测试函数上的寻优精度均值分别为0.236和3.129,均优于对比算法。在大规模仿真与实际边缘测试平台上,算法在保证任务截止时间满足率高于99%的同时,将系统综合运营成本降低35%,边缘节点内存利用率均值达67.54%,云边带宽利用率均值达54.56%,平均负载波动响应时间为0.77s,展现出对动态光电光电数据流负载的快速自适应能力。可见,该算法有效突破了传统深度强化学习算法的性能瓶颈,在电力云边协同光电光电数据流调度中实现了高可靠、低时延、低成本的调度目标,为电力光电感知网络与云端算力的融合提供了高效可行的决策支撑。

    Abstract:

    With the large-scale deployment of photoelectric sensing devices such as fiber-optic sensors and photoelectric transformers at the power edge side, massive high-bandwidth photoelectric sensing data streams pose severe challenges to the real-time transmission and processing of cloud-edge collaborative architectures. To address the online scheduling challenges of resource heterogeneity, complex task dependencies, and multi-objective conflicts in power cloud-edge collaborative data stream processing, this study models the scheduling process as a Markov decision process and proposes an imitation learning-guided deep deterministic policy gradient scheduling algorithm. The algorithm hierarchically integrates imitation learning with deep deterministic policy gradient and trains expert policies based on historical optimal scheduling logs to guide efficient exploration and stable optimization of the agent in highly dynamic photoelectric data stream scenarios. Meanwhile, a multi-objective collaborative objective function and a four-layer constraint system are constructed to achieve integrated collaborative decision-making for task offloading, memory allocation, and bandwidth reservation. The results show that the average optimization accuracies of the algorithm on Sphere and Rastrigin test functions are 0.236 and 3.129, respectively, both superior to the comparison algorithms. On large-scale simulation and actual edge testing platforms, the algorithm ensures a task deadline satisfaction rate exceeding 99% while reducing the overall system operating cost by 35%, achieving average edge node memory utilization of 67.54%, average cloud-edge bandwidth utilization of 54.56%, and an average load fluctuation response time of 0.77 s, demonstrating rapid adaptability to dynamic photoelectric data stream loads. It can be seen that this algorithm effectively breaks through the performance bottlenecks of traditional deep reinforcement learning algorithms, achieving high reliability, low latency, and low cost in the scheduling of power cloud-edge collaborative photoelectric data streams, providing efficient and feasible decision support for the integration of power photoelectric sensing networks and cloud computing power.

    参考文献
    相似文献
    引证文献
引用本文
分享
相关视频

文章指标
  • 点击次数:
  • 下载次数:
  • HTML阅读次数:
  • 引用次数:
历史
  • 收稿日期:2026-07-03
  • 最后修改日期:2026-07-03
  • 录用日期:2026-07-08
  • 在线发布日期:
  • 出版日期:
文章二维码