基于VMD和时空网络变分自编码器的负荷聚类

为了解决用户用电负荷曲线数据维度高、特征提取困难以及序列存在信号模态混叠的问题，本文提出使用变分模态分解（variational mode decomposition，VMD）和改进基于时空网络的变分自编码器（variational auto-encoders，VAE）对电力负荷曲线进行特征提取。通过模态分解得到信号的固有模态，对模态重构得到时序特征较明显的序列信号。再通过长短期记忆网络（long short-term memory network，LSTM）和卷积网络（convolutional neural networks，CNN）组成的时空变分自编码器进行潜在特征提取，并构建网络分类器来联合损失优化自编码器模型。最后使用Minibatchkmeans算法聚类并计算聚类中心。使用UCI数据集中葡萄牙居民用电量作为实验数据，通过实验结果表明经模态分解后通过降维再聚类的算法在戴维斯丁堡指数（Davies-Bouldin Index，DBI）和轮廓系数（Silhouette Coefficient，SC）上表现出较好效果。

Load Clustering Research Based on VMD and Spatiotemporal Network Variational Auto-encoder

In order to solve the problems of high data dimension, difficult feature extraction, and signal modal aliasing in the sequence, a feature extraction algorithm for power load curve based on variational mode decomposition(VMD) and improved spatiotemporal networks variational auto-encoder(VAE) was proposed in this paper. First of all, the intrinsic mode of the signal was obtained by modal decomposition, and the sequence signal with obvious timing characteristics was obtained by modal reconstruction. Then, the spatiotemporal networks variational auto-encoder composed of the long short-term memory network(LSTM) and the convolutional neural network(CNN) was used to extract potential features, and a network classifier was constructed to jointly associate the loss-optimized autoencoder model. Finally, the Minibatchkmeans algorithm was used to cluster and calculate the clustering center. Using the actual electricity consumption of Portuguese residents in the UCI dataset as the experimental data, The experimental results show that the algorithm of dimensionality reduction reclustering after modal decomposition has good results on the Davies-Bouldin Index (DBI) and Silhouette Coefficient (SC).