基于长短期记忆神经网络和改进型K-means聚类算法的居民峰谷时段划分模型  

A Residential Peak and Valley Time Division Model Based on Long Short-Term Memory and Improved K-Means Clustering Algorithm

作　　者：江兵[1] 李国荣 孙赵盟 庞宗强[1] JIANG Bing;LI Guorong;SUN Zhaomeng;PANG Zongqiang(School of Automation,Nanjing University of Posts and Telecommunications,Nanjing 210023,Jiangsu Province,China)

机构地区：[1]南京邮电大学自动化学院,江苏省南京市210023

出　　处：《现代电力》2021年第6期620-627,I0004,I0005,共10页Modern Electric Power

摘　　要：为了解决传统峰谷时段划分方法因只选取单一典型日而无法在较长时间范围内适用的问题,提出一种基于长短期记忆神经网络(long short-term memory,LSTM)和改进型K-means聚类算法的居民峰谷时段划分模型:首先对居民用户一整年的负荷数据进行有效性检查和归一化处理,保证数据的准确可靠;接着将处理后的负荷数据按照不同季节及不同日期类型进行相应的分类,保证分类的数据具有较强的相似性;然后将数据按分类分别加入LSTM进行训练,获得用户在不同分类下的负荷特征数据;最后利用改进型K-means聚类算法对训练得到的负荷特征数据进行聚类分析,并依据相应的权重矩阵及划分原则获得最终的时段划分结果。结果表明,相对于经典及当地的时段划分,所提方法的时段划分轮廓系数平均值更大,方差更小,更能反映居民用户实际的用电特点及用电规律,有利于挖掘用户侧需求响应潜力,获得更优的削峰填谷效果。To cope with the problem that traditional peak-valley time interval division method only selects single typical day so it cannot be applied to the situation of longer time range,based on long short-term memory(abbr.LSTM)neural network and improved K-means clustering algorithm a residential peak-valley time interval division model was proposed.Firstly,the validity check and normalization processing of the residential user’s load data of a whole year were performed to ensure that the data were accurate and reliable.Secondly,corresponding classification of the processed load data was implemented according to various seasons and different date type to ensure the classified data possessing stronger similarity.Thirdly,according to their classifications,the data was added to LSTM neural network for carrying on training to obtain load characteristic data of the user under different classifications.Finally,utilizing improved K-means clustering algorithm the clustering analysis on the load characteristic data of the user,which was obtained from the training,was carried out,and according to corresponding weight matrix and division principle the final time interval division result was gained.Simulation results show that in comparison with the classic and local time interval divisions,the average value of time interval division silhouette coefficient obtained by the proposed method is larger,meanwhile the variance is smaller,so the proposed method can reflect both actual characteristics and pattern of residential user’s electricity consumption better,and is in favour of exploiting user-side demand response potentialities as well as obtaining a better peak-load shifting and valley-filling effect.

关 键 词：峰谷时段划分 长短期记忆神经网络 改进型Kmeans算法 聚类分析 轮廓系数 

分 类 号：TM715[电气工程—电力系统及自动化]