A Multiple Instance Regression Approach to Electrical Load Disaggregation

Non-Intrusive Load Monitoring (NILM) provides detailed information on the consumption of individual appliances in a building and represents an effective method to reduce the electricity consumed in the residential sector. Supervised Deep Learning approaches have achieved the state-of-the-art for NILM but require knowledge of strongly labeled data, i.e., annotated at the sample level. This data is costly to obtain since it requires multiple sensors to measure electrical quantities and the involvement of the end-users. This work proposes a Multiple Instance Regression approach to NILM using a Convolutional Recurrent Neural Network (CRNN) to reduce the amount of strongly labeled data required for training and improve performance. Instances of strongly labeled data are here represented by raw samples of active power and are aggregated into bags containing weak information rep-resented by the average power consumption in a bag. Using this information, the network is trained to disaggregate appliances' power profiles with sample resolution. The results obtained on the UK-DALE dataset demonstrated the approach's effectiveness in reducing the labeling cost and improving the performance: the average Mean Absolute Error reduces by 3.06 W when weak information is used in the CRNN and by 8.88 W compared to the Sequence-to-Point method.