Realizing multifractality of smart meter data for household characteristic prediction

Driven by the efficient and personalized services and better management of the distribution networks, power utilities are always willing to recognize the household characteristics which reflect the dwelling and social-demographic characteristics of customers. As a key component of the distribution grid monitoring systems, smart meters installed at each customer's household offer a non-intrusive approach to predict the household characteristics. The resolution of energy consumption data has a great impact on prediction accuracy as fine granularity data can better characterize the household energy usage profile features. However, collecting high-resolution energy consumption data may not be feasible for power utilities due to the substantial cost of upgrading the existing Advanced Metering Infrastructure (AMI). This paper aims to propose a cost-effective framework for the accurate prediction of household characteristics by realizing the multifractal characteristics of residential electric energy consumption data recorded from smart meters. First, Multiscale Multivariate Multifractal Detrended Fluctuation Analysis (MMV-MFDFA) is performed on the energy consumption data to analyze the multiscale relationships of the residential energy usage profile and reveal the multifractal structures embedded in the energy data. Then, Self-Adaptive Multifractal Interpolation (SAMI) is developed to reconstruct quasi high-resolution energy consumption data from which distinctive features relevant to household characteristics are extracted by Mathematical Morphological Decomposition (MMD). Such features are further integrated with an advanced deep learning algorithm - Deep Forest (DF) for household characteristic prediction. The proposed method is validated using real-life smart meter data from Ireland.