Recognition and classification of typical load profiles in buildings with non-intrusive learning approach

The recent increasing spread of Advanced Metering Infrastructure (AMI) has enabled the collection of a huge amount of building related-data which can be exploited by both energy suppliers and users to gain insight on energy consumption patterns. In this context, data analytics-based methodologies can play a key role for performing advanced characterization, benchmarking and classification of buildings according to their typical energy use in the time domain. Traditionally, energy customers are classified according to their building end-use category. However, buildings belonging to the same category can exhibit very different energy patterns making ineffective this kind of a-priori categorization. For this reason, load profiling frameworks have been developed in the last decade to identify homogenous groups of buildings with similar daily energy profiles. The present study proposes a non-intrusive customer classification process, which does not use as predictive attributes in-field load monitoring data for the classification of unknown customers, but rather monthly energy bills and additional information on customers' habits collected by means of a phone survey. The proposed classification process is developed by analysing hourly energy consumption data of 114 electrical customers of an Italian Energy Provider. The representative daily load profiles are grouped using the Follow the Leader clustering algorithm and a globally optimal decision tree is employed to build a supervised classification model. The model, compared to a baseline recursive partitioning tree, leads to an increase of accuracy of about 6%. Eventually, the procedure exploits energy bill data also for estimating the magnitude of typical load profiles.