Self-training convolutional autoencoder for consumer characteristics identification with imbalance datasets

Abstract

Consumer characteristics can help energy utilities implement efficient demand response programs and personalized services. However, there are two problems in obtaining consumer characteristics from smart meter data: label scarcity and class imbalance. The two problems greatly reduce the effectiveness of the existing supervised learning classifiers. Therefore, this paper proposes a self-training convolutional autoencoder (STCAE) framework for consumer characteristics identification with imbalance datasets. STCAE has three modules including encoder, decoder, and classifier. For the label scarcity problem, a seasonal pretext is designed to enable learning unlabeled samples through encoder–decoder structure of STCAE. For the class imbalance problem, a piecewise synthetic minority over-sampling technique (PSMOTE) method is designed for class rebalancing. Case studies and comparison studies demonstrate the validity of each major component and the superiority of STCAE over existing methods. In comparison studies of twelve consumer characteristics identification, the average improvements of accuracy and Matthews correlation coefficient are higher than 6.32% and 35.18%, respectively. Furthermore, in the ablation studies, STCAE performs better than the framework with different components in most consumer characteristics.