一种用于非侵入式负荷分解的改进时域卷积网络
An Improved Time-Domain Convolutional Network for Non-Invasive Load Decomposition

非侵入式负载监测(Non-Intrusive Load Monitoring)是通过监测用户电表的总用电量数据并应用负载分解算法，以获取各个电器实时用电量数据。然而，非侵入式负载分解作为一种单通道盲源分离问题，由于其固有的无法识别性导致实现分解困难。深度学习逐渐成为解决NILM问题的主流方法，得益于可用的数据、计算能力和深度网络训练算法模型。其中，序列到点(seq2point)负荷分解模型(Zhang)实现了最先进的预测效果。利用TCN网络架构，使用双向扩张卷积结构代替因果卷积结构，扩大了网络的感受野，并使用非线性激活函数GELU和层标准化(LN)方法，通过残差连接的思想搭建了时域卷积网络模型。最后，在公共数据集UK-DALE上对算法进行了测试，并选择平均绝对误差和F-score作为主要评价指标来评估算法的性能。通过对比两种算法在相同周期数据集上的分解结果，发现基于TCN改进的模型，显著提高了负荷分解的性能指标。
Non-Intrusive Load Monitoring (NILM) involves monitoring the total electricity consumption data from the customer’s meter and applying a load disaggregation algorithm to obtain real-time electricity consumption data for individual appliances. However, non-intrusive load disaggregation, as a single-channel blind source separation problem, is challenging to achieve due to its inherent unidentifiability. Deep learning is gradually becoming a mainstream approach for solving the NILM problem, thanks to the availability of data, computational power, and advanced deep network training algorithms. Among them, the Sequence-to-Point (Seq2Point) load decomposition model (Zhang) achieves state-of-the-art prediction results. The TCN network architecture is utilized to expand the sensory field of the network by employing a bidirectional dilation convolution structure instead of a causal convolution structure. A time-domain convolutional network model is constructed based on the concept of residual connectivity using a nonlinear activation function, GELU, and a layer normalization (LN) method. Finally, the algorithms were tested on the public dataset UK-DALE. The mean absolute error and F-score were selected as the primary evaluation metrics to assess the algorithms’ performance. By comparing the decomposition results of the two algorithms on the same dataset period, we found that the enhanced model based on TCN significantly improves the performance metrics of load decomposition.