Lightweight Non-Intrusive Load Monitoring Employing Pruned Sequence-to-Point Learning
Jack R. Barber, Heriberto Cuayáhuitl, Mingjun Zhong, Wenpeng Luan
Abstract
Non-intrusive load monitoring (NILM) is the process in which a household's total power consumption is used to determine the power consumption of household appliances. Previous work has shown that sequence-to-point (seq2point) learning is one of the most promising methods for tackling NILM. This process uses a sequence of aggregate power data to map a target appliance's power consumption at the midpoint of that window of power data. However, models produced using this method contain upwards of thirty million weights, meaning that the models require large volumes of resources to perform disaggregation. This paper addresses this problem by pruning the weights learned by such a model, which results in a lightweight NILM algorithm for the purpose of being deployed on mobile devices such as smart meters. The pruned seq2point learning algorithm was applied to the REFIT data, experimentally showing that the performance was retained comparing to the original seq2point learning whilst the number of weights was reduced by 87%. Code:https://github.com/JackBarber98/pruned-nilm