Estimation of household characteristics with uncertainties from smart meter data

The knowledge of household characteristics can help energy providers carry out more personalized demand-side management programs. Obtaining such information through surveys is costly and time-consuming in practice. This paper proposes a novel estimation method for household characteristics with uncertainties using the residential electricity consumption data. To alleviate the class imbalance problem in the dataset, a dynamic time warping sampling (DTWS) method is proposed to generate synthetic data for the minority class. To overcome the problem that the existing methods for identifying household characteristics cannot provide the confidence level of the results, a Bayesian convolutional neural network (BCNN) model is developed for feature extraction and characteristic identification with uncertainties. These quantified uncertainties can be regarded as a measure of confidence and can be used to target customers more effectively for energy efficiency and demand response programs. The effectiveness of the proposed model is validated by experiments on ground truth data.