What's Up for theWeekend? Exploiting Day Type Information in Non-Intrusive Load Monitoring

Interactions with most electrical appliances in buildings are governed by temporal patterns. Daily routines and activities have an inherent impact on (1) the types of appliances being used, (2) at what time they are operated, and (3) for what duration. Similarly, most commercial and industrial energy consumers exhibit consumption characteristics that depend on time and the day of the week. Almost all published load disaggregation methods are, however, agnostic to the time and day during which their input data has been collected. In this work, we hence study whether an improved load disaggregation is possible when input data are partitioned by the day type, and separate models are created for each partition. We specifically consider two methods to partition daily energy consumption into two sets. The first method is based on separating weekends and holidays (from weekdays), while the second method uses.. -means clustering to autonomously separate the data into clusters. We evaluate the disaggregation performance of each clustering method using three datasets that were collected in different settings. Our results show that an improvement up to 40 % can be reached in the disaggregation accuracy, as measured by the MAE, when temporal information is considered in the disaggregation process.