基于多模态动态融合与细粒度安全检索的电力调度智能问答双驱模型设计
DOI:
CSTR:
作者:
作者单位:

中国南方电力调度控制中心

作者简介:

通讯作者:

中图分类号:

基金项目:

基于大数据和人工智能的调度智库与智写系统技术研究与开发科技项目(ZDKJXM20240064)


Design of a Dual-Driven Intelligent Question-Answering Model for Power Dispatch Based on Multimodal Dynamic Fusion and Fine-Grained Security Retrieval
Author:
Affiliation:

China Southern Power Dispatching Control Center

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:

    针对含光电监测链路、光纤通信量测与智能传感终端运行信息的电力调度智能问答中,安全约束难以穿透至语义层、动态知识更新滞后等问题,提出一种双驱安全增强型语义融合与问答生成架构。该架构通过双重注意力机制增强的轻量语义编码模块,实现文本、光电量测、设备状态数据与电网拓扑数据的多模态动态融合,提升复杂调度场景下的语义理解能力;通过集成细粒度检索的生成模块,实现安全规程与实时知识的高精度匹配和动态注入。两模块基于双路注意力门控与闭环知识流协同机制,实现安全规则与实时工况的深度耦合。实验结果表明,该方法语义一致性均值达0.91,领域知识准确率持续高于0.92,响应时延稳定在15.3 ms,知识检索召回率达91.8%,单次能耗为24.8 mAh,内存占比优化至47.5%,整体优于对比方法。实际测试中,该方法在接地短路故障场景下保持18 ms低时延,在新能源波动场景下能耗低至24.1 mAh。综上,所提架构有效提升了调度系统在复杂运行环境下的认知深度、响应实时性与安全合规性,可为高比例新能源电网源网荷储协同调度及光电监测数据驱动的运行状态认知提供智能问答支撑。

    Abstract:

    In intelligent question-answering systems for power dispatching that involve optoelectronic monitoring links, optical fiber communication measurements, and operational information from intelligent sensing terminals, challenges such as the difficulty in penetrating security constraints into the semantic layer and the lag in dynamic knowledge updates are addressed. A dual-driven security-enhanced semantic fusion and question-answering generation architecture is proposed. This architecture achieves multimodal dynamic fusion of text, photovoltaic measurement, equipment status data, and grid topology data through a lightweight semantic encoding module enhanced by a dual attention mechanism, thereby enhancing semantic understanding capabilities in complex dispatching scenarios. By integrating a fine-grained retrieval generation module, it achieves high-precision matching and dynamic injection of safety procedures and real-time knowledge. The two modules, based on a dual-path attention gating and closed-loop knowledge flow coordination mechanism, enable deep coupling of safety rules and real-time operating conditions. Experimental results show that the method achieves a semantic consistency mean of 0.91, with domain knowledge accuracy consistently exceeding 0.92, response latency stabilizing at 15.3 ms, knowledge retrieval recall rate reaching 91.8%, single energy consumption of 24.8 mAh, and memory usage optimized to 47.5%. Overall, this method outperforms comparative approaches. In practical testing, the method maintains a low latency of 18 ms in grounding short-circuit fault scenarios and achieves energy consumption as low as 24.1 mAh in renewable energy fluctuation scenarios. In summary, the proposed architecture effectively enhances the cognitive depth, real-time response, and safety compliance of dispatching systems in complex operating environments. It can provide intelligent question-answering support for collaborative dispatching of source-grid-load-storage in high-proportion renewable energy grids and operational state cognition driven by optoelectronic monitoring data.

    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