Scholarlink

With the ongoing digitalization and intelligent transformation of power systems, the grid data platform, as a critical data infrastructure, has a direct impact on the operational continuity and regulation security of power systems. However, traditional anomaly detection methods struggle to address the high-dimensional, multi-source heterogeneous, and dynamically evolving characteristics of platform operational data, necessitating intelligent prediction approaches with global modeling and temporal perception capabilities. To address this challenge, this study proposes an anomaly trend prediction model integrating a Transformer encoder and BiLSTM decoder architecture. The model combines multi-head attention mechanisms with bidirectional sequence modeling to achieve high-precision trend forecasting for multi-indicator system states. Additionally, we develop a comprehensive system implementation framework supporting batch modeling and scheduled training for multiple subsystem tasks, demonstrating strong engineering adaptability and scalability. Experimental results show that the proposed model achieves a coefficient of determination ( \(\text {R}^2\) ) of 0.9003 on the test set, exhibiting superior early-warning capability and cross-task adaptability. The model has been deployed on a provincial-level power data platform, enabling joint prediction of key indicators across subsystems including OGG, DataWorks, and MaxCompute. This work provides an efficient solution for intelligent monitoring and operation management of power system platforms, effectively balancing global attention modeling and temporal continuity.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Anomaly Prediction in Power Data Platforms Using A Transformer-BiLSTM Model

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ruixuan Lu,
  • Zhaohong Jia,
  • Shichuan Ding,
  • Jin Zhang,
  • Zhen Yao,
  • Zhenhua Huang

摘要

With the ongoing digitalization and intelligent transformation of power systems, the grid data platform, as a critical data infrastructure, has a direct impact on the operational continuity and regulation security of power systems. However, traditional anomaly detection methods struggle to address the high-dimensional, multi-source heterogeneous, and dynamically evolving characteristics of platform operational data, necessitating intelligent prediction approaches with global modeling and temporal perception capabilities. To address this challenge, this study proposes an anomaly trend prediction model integrating a Transformer encoder and BiLSTM decoder architecture. The model combines multi-head attention mechanisms with bidirectional sequence modeling to achieve high-precision trend forecasting for multi-indicator system states. Additionally, we develop a comprehensive system implementation framework supporting batch modeling and scheduled training for multiple subsystem tasks, demonstrating strong engineering adaptability and scalability. Experimental results show that the proposed model achieves a coefficient of determination ( \(\text {R}^2\) ) of 0.9003 on the test set, exhibiting superior early-warning capability and cross-task adaptability. The model has been deployed on a provincial-level power data platform, enabling joint prediction of key indicators across subsystems including OGG, DataWorks, and MaxCompute. This work provides an efficient solution for intelligent monitoring and operation management of power system platforms, effectively balancing global attention modeling and temporal continuity.