Scholarlink

<p>Non-intrusive load monitoring (NILM) is pivotal in optimizing energy management and enhancing efficiency by analyzing the energy consumption patterns of individual appliances within a household. However, the deployment of NILM algorithms on edge devices is impeded by hardware resource limitations, as traditional deep learning models are computationally intensive and thus ill-suited for resource-constrained edge environments. This study proposes an Integrated Pruning and Distillation-based Non-Intrusive Load Monitoring (IPD-NILM) multi-model joint compression method, which combines layer-wise iterative pruning with knowledge distillation techniques to achieve efficient model compression while maintaining high performance and stability. Through the utilization of model reconstruction techniques and deployment procedures, the IPD-NILM methodology enables the transfer of intricate models from high-performance platforms to edge devices. The results from the UK-DALE and REDD datasets show that IPD-NILM achieves a 72% reduction in the number of trainable parameters and a 63% decrease in inference time, all while retaining performance parity, successfully achieving the practical application of NILM technology.</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

Data Model Compression Method Based on Iterative Pruning and Distillation for Edge Non-Intrusive Load Monitoring

  • 期刊论文

  • 发表日期：

    2025年7月28日

查看全文
  • Sheng Zhao,
  • Shukang Li,
  • Ping Gao,
  • Shiwei Ge,
  • Zhengjiang Zhang,
  • Liang Shu

摘要

Non-intrusive load monitoring (NILM) is pivotal in optimizing energy management and enhancing efficiency by analyzing the energy consumption patterns of individual appliances within a household. However, the deployment of NILM algorithms on edge devices is impeded by hardware resource limitations, as traditional deep learning models are computationally intensive and thus ill-suited for resource-constrained edge environments. This study proposes an Integrated Pruning and Distillation-based Non-Intrusive Load Monitoring (IPD-NILM) multi-model joint compression method, which combines layer-wise iterative pruning with knowledge distillation techniques to achieve efficient model compression while maintaining high performance and stability. Through the utilization of model reconstruction techniques and deployment procedures, the IPD-NILM methodology enables the transfer of intricate models from high-performance platforms to edge devices. The results from the UK-DALE and REDD datasets show that IPD-NILM achieves a 72% reduction in the number of trainable parameters and a 63% decrease in inference time, all while retaining performance parity, successfully achieving the practical application of NILM technology.