Data Augmentation and Class Based Model Evaluation for Load Disaggregation Based on Deep Learning

Load disaggregation aims to disaggregate the power usage of individual appliance based only on aggregated power monitoring data. Applications such as appliance monitoring, appliance-based power metering can be realized based on this technique. In recent years, research has been done to apply deep learning to load disaggregation and initial results can be found in literature. However, for all machine learning based models, the class imbalance problem or the skewed distribution of values in training datasets, has long been recognized as a severe problem dragging down the overall accuracy of the model trained. This is also a common problem in current public datasets for load disaggregation research due to natural imbalance of appliance on/off time. In this work, we address the data imbalance problem by proposing a novel data augmentation method that combines both original data sequences and selected data sequences. Moreover, we propose a more fair and natural evaluation method to compare the models’ performance under imbalance datasets. Extensive empirical study shows that the proposed method can achieve 53.48–83.81% performance enhancement over the current state-of-the-art model.