考虑数据不均衡的居民用户负荷曲线分类方法

由于用户用电行为的多样性和随机性，负荷数据存在分布不均衡的问题，传统负荷曲线分类方法在处理不均衡数据时性能较差，为此，提出一种改进K-means与LSTM（long short term memory network）-CNN（convolutional neural network）分类模型结合的负荷曲线分类方法。首先，为提升K-means算法对不均衡数据的聚类效果，基于密度峰值聚类算法（density peaks clustering，DPC）思想，提出一种相对k近邻密度峰值初始聚类中心选取方法（related k-nearest neighbor density peaks，RKDP），将其作为K-means算法的初始中心进行聚类；其次，为提高RKDP_K-means处理高维负荷数据的性能，采用LSTM自编码器进行特征降维后再聚类（LSTM-auto-encoder RKDP_K-means，LARK）获得精准类别标签；最后，基于LSTM和CNN网络分别提取负荷特征构建负荷曲线分类模型，实现对大规模负荷曲线的分类。算例表明在大规模负荷曲线分类时，相比于LARK算法，本文所提方法轮廓系数指标提升29.7%，效率提升3.46倍，具有良好的负荷曲线分类效果。

Residential user load curve classification method considering data imbalance

With the development of smart grids, the dimension and volume of load data continue to increase. At the same time, due to the diversity and randomness of users'' electricity consumption behaviors, the imbalance of load data classes is increasingly obvious. Traditional load curve classification technologies have become ineffective to deal with the im-balanced class problem of data. Therefore, an algorithm combing improved K-means with long short term memory net-work (LSTM) and convolutional neural network (CNN) classification model is proposed. Firstly, to improve the classifica-tion accuracy of the K-means on imbalanced data, a method of relative k-nearest neighbor density peaks(RKDP) is proposed to select the initial clustering centre of K-means. Secondly, an auto-encoder based on LSTM (LSTM-AE) is used to extract load characteristics from high dimensional data, and combined with RKDP_K-means to obtain accu-rate load profiles labels. Finally, a classification model based on the LSTM and CNN networks is used to realize the classification of large-scale load profiles. The results show that compared with LARK(LSTM-AE RKDP_K-means) al-gorithm, the Silhouette Coefficient of the proposed method is improved by 29.7%, and the efficiency is improved by 3.46 times.