Stochastic generation of residential load profiles with realistic variability based on wavelet-decomposed smart meter data
Robbert Claeys, Rémy Cleenwerck, Jos Knockaert, Jan Desmet
Abstract
Residential smart meter data with high time resolution are integral to many data-driven applications, ranging from hosting capacity studies to R&D activities of private enterprises. However, privacy legislation restricts public availability of large-scale datasets. Furthermore, existing datasets may suffer from imbalances in terms of underrepresented classes. To address these concerns, this study presents a novel decomposition–recombination approach for generating synthetic load profiles that exhibit realistic variability and demand peaks. High-frequency load profiles are decomposed into a low-frequency base load and high-frequency variability at the daily level through a discrete wavelet transformation. Components from different households are subsequently rescaled, shifted and recombined in a stochastic load profile generator to obtain new daily load profiles with high-fidelity behavior. The performance of this generator is evaluated through benchmarking, resulting in a mean average error of 0.09 kW on an average value of less than 3 kW for the daily peaks, whilst preserving their seasonality. The introduced load profile generator is validated as an alternative to privacy-sensitive residential smart meter data in a hosting capacity case study. The analysis focuses on the voltage drop caused by residential electric vehicle charging, considering both real and synthetic data. The synthetic data demonstrated voltage drops with a mean average error less than 0.2 V for the 10th and 90th percentile when benchmarked with respect to the real voltage level distribution. The introduced decomposition–recombination method is shown to accurately capture the high-frequency variability and peak behavior, and is suitable for practical applications at the daily level.