| A household profile is a description of a customer that includes demographic,geographic,psychographic characteristics,and purchase history as well as other personalized dimensions.In this paper,we focus on the household profiles relating to electricity use behaviors,which mainly cover four categories includin g dwelling characteristics,socio-demographic,appliances and heating and attitudes towards energy.Deeper insight into household profiles is of great significance for utilities,demand response aggregators,government departments etc.to flexibly gather and adjust demand-side resources,and implement personalized green energy-saving and efficiency-enhancing projects,which will help households to improve the overall perception of energy efficiency level and the degree of energy-saving response,thus promoting the realization of China’s "30·60" dual carbon goal.Based on smart meter data,this paper focuses on the problem of analyzing specific household profiles.The main contents are as follows:(1)The feature extraction and selection of smart meter data.The research essence of household profiles identification is a label classification problem.The quality of extracted features has a direct impact on the performance upper bound of the classification model.However,the existing researches about feature extraction are only concentrated in a single domain and the potential patterns contained in smart meter data can’t be comprehensively analyzed from multiple angles.To this end,a time-frequency domain feature combination method based on smart meter data is proposed to identify household profiles.Firstly,considering the differences of households’ electricity consumption levels,a lot of time-domain features are extracted(consumption features,statistical features,etc.).On this basis,discrete wavelet transform is used to decompose the initial average daily power curve of each household and the frequency-domain features that reflect the difference of household load fluctuation patterns are extracted.Then,the random forest feature selection algorithm is applied to sort the importance of extracted features and select the significant determinants to reduce information redundancy.Compared with the case of only time-domain features or only-frequency domain,simulation results show that the combination of time and frequency domain features effectively improves the accuracy and the performance of the household profiles identification model.(2)A household profiles identification method based on semi supervised learning.The existing household profile identification methods are all supervised learning-based methods.Such methods can achieve promising performance in the case of sufficient labeled data but show low accuracy if labelled data is insufficient or even unavailable.However,the acquisition of accurate labelled data(usually obtained by survey)is very difficult,costly,and time-consuming in practice due to various reasons.To this end,a semi-supervised learning method is applied into household profile identification to improve the identification performance,which not only classifies the labeled samples well but also makes full use of the potential distribution information covered by the substantial unlabeled samples.Simulation results show that the proposed semi-supervised learning approach outperforms the supervised learning methods in the case of limited labeled samples.What’s more,when the proportion of labeled samples trained by supervised learning methods is set to 10 times of semi-supervised learning methods,the semi-supervised methods still show strong advantages over other supervised learning methods,which shows that the proposed method can effectively reduce the labeling cost. |
PDF Full Text Request