Scholarlink

<p>Existing non-intrusive load monitoring methods often have difficulty in effectively dealing with the challenges of dynamic load fluctuations, limited generalization ability under small sample conditions, and inefficient deployment on resource-constrained edge devices. To compensate for these deficiencies, a hybrid framework based on a hidden Markov model and a graph convolutional network is proposed in this study. It realizes adaptive feature extraction in the unsteady power grid environment by introducing a dynamically updatable electrical apparatus association graph. Unlike previous hybrid learning methods or meta-learning methods, which rely on fixed topology or require a large number of parameter adjustments, the proposed edge-oriented model compression strategy can effectively control the memory overhead while ensuring real-time performance. The experimental results indicate that the proposed framework performs well on both SynNILM-2024 and EDF-LPWA datasets. Specifically, 93.4% topology reconstruction accuracy is achieved in the dynamic topology test, which is 17.8 percentage points higher than the existing optimal methods. Besides, in the small sample learning task, only 15 samples can be used to control the new appliance recognition error to 4. 1 ± 0.7%. It reduces the data requirement by about 81.3% compared to the typical method. On the Raspberry Pi 4B platform, the inference delay is only 17.9&#xa0;ms. Besides, the memory footprint is compressed to 312&#xa0;MB, which is better than the existing edge deployment scheme. Therefore, this study provides a scalable solution to drive real-time load monitoring towards a more resilient and energy-aware smart grid.</p>

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

Improved hidden Markov-graph convolutional network energy decomposition algorithm for non-invasive load monitoring

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月27日

查看全文
  • Yangrui Zhang,
  • Chao Zhang,
  • Yakui Zhu,
  • Bingyu Zhang,
  • Hongying Wang,
  • Shaokang Feng

摘要

Existing non-intrusive load monitoring methods often have difficulty in effectively dealing with the challenges of dynamic load fluctuations, limited generalization ability under small sample conditions, and inefficient deployment on resource-constrained edge devices. To compensate for these deficiencies, a hybrid framework based on a hidden Markov model and a graph convolutional network is proposed in this study. It realizes adaptive feature extraction in the unsteady power grid environment by introducing a dynamically updatable electrical apparatus association graph. Unlike previous hybrid learning methods or meta-learning methods, which rely on fixed topology or require a large number of parameter adjustments, the proposed edge-oriented model compression strategy can effectively control the memory overhead while ensuring real-time performance. The experimental results indicate that the proposed framework performs well on both SynNILM-2024 and EDF-LPWA datasets. Specifically, 93.4% topology reconstruction accuracy is achieved in the dynamic topology test, which is 17.8 percentage points higher than the existing optimal methods. Besides, in the small sample learning task, only 15 samples can be used to control the new appliance recognition error to 4. 1 ± 0.7%. It reduces the data requirement by about 81.3% compared to the typical method. On the Raspberry Pi 4B platform, the inference delay is only 17.9 ms. Besides, the memory footprint is compressed to 312 MB, which is better than the existing edge deployment scheme. Therefore, this study provides a scalable solution to drive real-time load monitoring towards a more resilient and energy-aware smart grid.