基于深度自编码器的分钟级负荷数据聚类分析

Cluster Analysis of Minute Load Data Based on Deep Autoencoder

在数字化电网的全面建设和电力市场持续改革的背景下,电力服务商积极开展以负荷数据聚类分析为基础的用电行为分析。为了深入分析用户的用电行为模式,提出基于深度自编码器的分钟级负荷数据聚类分析方法。首先基于信息熵重构负荷数据,保留负荷数据的形态特征和提高数据的可区分性;接着提出深度自编码器的特征提取方法,同时利用边界少数样本过采样算法生成新的训练样本,对深度自编码器网络模型进行两阶段训练;最后基于欧式距离和动态时间扭曲距离的双尺度距离,计算负荷数据特征的相似性,以双尺度距离作为K-means算法的输入数据得到负荷聚类结果。基于南京市某台区的分钟级负荷数据的算例分析表明,所提方法提高了不同负荷数据分类的准确性。

Under the background of all-round construction of digital power grid and continuous reform of power market,power service providers actively carry out power consumption behavior analysis based on cluster analysis of load data.In order to deeply analyze users'electricity consumption behavior patterns,this paper proposes a minute-level load data clustering analysis method based on deep autoencoder.Firstly,the load data is reconstructed based on information entropy,the morphological characteristics of the load data are retained and the data distinguishability is improved.Then,the feature extraction method of deep autoencoder is proposed.At the same time,a new training sample is generated by using the boundary minority sample oversampling algorithm,and the network model of deep autoencoder is trained in two stages.Finally,based on the double-scale distance of Euclidean distance and dynamic time warping distance,the similarity of load data features is calculated,and the double-scale distance is used as the input data of K-means algorithm to get the load clustering results.An example analysis based on the minute load data of a station in Nanjing shows that this method improves the accuracy of different load data classification.