Scholarlink

<p>Deep learning technology can realize the rapid identification of non-intrusive loads. Still, it relies on a large number of offline labeled samples for model training, which increases the cost of the modeling. Due to the randomization behavior of the daily electricity consumption and the sparsity of the unknown load sampled data, using the new unknown load sample for modeling results in low accuracy and poor reliability. These issues significantly hinder the promotion and use of non-intrusive load identification (NILI) technology within the electric power industry. In this paper, we present a load sparse sample recognition framework based on pre-training, KG-DBiLSTM, which enhances the fusion feature representation ability and improves the accuracy of small sample load identification through reliable knowledge transfer. KG-DBiLSTM uses the adaptive Holt-Winter (Ad-HW) algorithm to detect and denoise the raw load data, improving data availability. To learn more key feature information of load sampled data and to enhance the representation interpretability, we propose a dual-layer Transformer network, named Du-Transformer, which improves the fusion features extraction ability in the pre-training stage. Load knowledge is introduced into feature representation learning, accurately capturing the bidirectional dependence of the load time series sampled data. Then, the load fusion features are used to train a DBiLSTM network as the load type recognizer, effectively reducing the error rate of the model recognition. To ensure the reliability of the model parameters fine-tuning and transfer, we present a weighted cosine similarity (WCS) measurement method; WCS can eliminate the difference between the load feature fusion and parameters transfer in the fine-tuning stage. A large number of experiments on the REDD, UK-DALE, and RDSA datasets reveal that our method has a higher unknown load identification accuracy and good stability. In cases where accuracy is met, a small sample set is used to complete the fine-grained load identification, and this method has strong technical applicability in the 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

Knowledge fusion guided pre-training non-intrusive load sparse sample recognition

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Yanmei Jiang,
  • Lin Liu,
  • Mingsheng Liu,
  • Hao Peng,
  • Jianhua Li,
  • Shiyuan Yu,
  • Xiong Li

摘要

Deep learning technology can realize the rapid identification of non-intrusive loads. Still, it relies on a large number of offline labeled samples for model training, which increases the cost of the modeling. Due to the randomization behavior of the daily electricity consumption and the sparsity of the unknown load sampled data, using the new unknown load sample for modeling results in low accuracy and poor reliability. These issues significantly hinder the promotion and use of non-intrusive load identification (NILI) technology within the electric power industry. In this paper, we present a load sparse sample recognition framework based on pre-training, KG-DBiLSTM, which enhances the fusion feature representation ability and improves the accuracy of small sample load identification through reliable knowledge transfer. KG-DBiLSTM uses the adaptive Holt-Winter (Ad-HW) algorithm to detect and denoise the raw load data, improving data availability. To learn more key feature information of load sampled data and to enhance the representation interpretability, we propose a dual-layer Transformer network, named Du-Transformer, which improves the fusion features extraction ability in the pre-training stage. Load knowledge is introduced into feature representation learning, accurately capturing the bidirectional dependence of the load time series sampled data. Then, the load fusion features are used to train a DBiLSTM network as the load type recognizer, effectively reducing the error rate of the model recognition. To ensure the reliability of the model parameters fine-tuning and transfer, we present a weighted cosine similarity (WCS) measurement method; WCS can eliminate the difference between the load feature fusion and parameters transfer in the fine-tuning stage. A large number of experiments on the REDD, UK-DALE, and RDSA datasets reveal that our method has a higher unknown load identification accuracy and good stability. In cases where accuracy is met, a small sample set is used to complete the fine-grained load identification, and this method has strong technical applicability in the smart grid.