|
| 提升非侵入式负荷辨识准确度的改进型自编码器 |
| Accuracy improved denoising autoencoder for non-intrusive load monitoring |
| 投稿时间：2020-12-01 |
| DOI：10.3969/j.issn.1005-5630.2021.02.002 |
| 中文关键词: 非侵入式负荷辨识 电器能耗分解 自编码器 神经网络 |
| 英文关键词:non-intrusive load monitoring electric applianc energy disaggregation autoencoder neural network |
| 基金项目:国家自然科学基金(61074087、 61703277、 11502145) |
|
| 摘要点击次数: 2517 |
| 全文下载次数: 1726 |
| 中文摘要: |
| 非侵入式负荷辨识能够实现电器能耗监测、提高能源使用效率。针对电器混合能耗分解的任务，提出了一种包含改进型去噪自编码器的神经网络方法，用于分解低频采样的混合功率。该方法首先对不同电器分别训练一个卷积去噪自编码器神经网络，然后使用滑动窗口的方式将原始电表采样功率数据通过相应神经网络逐一进行局部分解，最后合成目标电器的干净时序功率数据。研究表明，在REDD数据集上进行实验时，该方法对电器负荷分解准确度较现有方法提高至少12%。准确度的提升归因于本方法在网络结构上加入了卷积层、批归一化和修正线性单元，并在样本预处理上零值化所有半运行训练样本标签，促使神经网络能充分地利用其有限的容量。 |
| 英文摘要: |
| Non-intrusive load monitoring facilitates the monitoring of electric appliance energy usage and the improvement on energy efficiency. Focusing on the task of disaggregating mixed energy consumption, this paper proposes a method that involves improved denoising autoencoder neural network to disaggregate mixed power sampled at low frequency. This method firstly trains a convolutional denoising autoencoder network for every electric appliance. Subsequently, the sliding window method is employed to disaggregate the mixed power segments, which are later combined to form the clean complete disaggregated time series power data. The research shows that the experiment on the REDD dataset demonstrates at least 12% improvement on disaggregation accuracy compared with current methods; and this improvement in performance attributes to the structural change that adds convolutional layers, batch normalization and rectified linear unit, as well as the sample preprocessing that innovatively zeros out labels of samples containing incomplete active sections forcing the network to better exploit its limited capacity. |
| HTML 查看全文 查看/发表评论 下载PDF阅读器 |
| 关闭 |
