A clustering approach to domestic electricity load profile characterisation using smart metering data

摘要

The availability of increasing amounts of data to electricity utilities through the implementation of domestic smart metering campaigns has meant that traditional ways of analysing meter reading information such as descriptive statistics has become increasingly difficult. Key characteristic information to the data is often lost, particularly when averaging or aggregation processes are applied. Therefore, other methods of analysing data need to be used so that this information is not lost. One such method which lends itself to analysing large amounts of information is data mining. This allows for the data to be segmented before such aggregation processes are applied. Moreover, segmentation allows for dimension reduction thus enabling easier manipulation of the data. Clustering methods have been used in the electricity industry for some time. However, their use at a domestic level has been somewhat limited to date. This paper investigates three of the most widely used unsupervised clustering methods: k-means, k-medoid and Self Organising Maps (SOM). The best performing technique is then evaluated in order to segment individual households into clusters based on their pattern of electricity use across the day. The process is repeated for each day over a six month period in order to characterise the diurnal, intra-daily and seasonal variations of domestic electricity demand. Based on these results a series of Profile Classes (PC's) are presented that represent common patterns of electricity use within the home. Finally, each PC is linked to household characteristics by applying a multi-nominal logistic regression to the data. As a result, households and the manner with which they use electricity in the home can be characterised based on individual customer attributes.