Generating realistic load profiles in smart grids: An approach based on nonlinear independent component estimation (NICE) and convolutional layers

The utilization of energy consumption data is crucial for efficient operation and planning in smart grids. Nonetheless, certain obstacles need to be addressed, such as high computational costs, data security and privacy concerns, and significant expenses associated with installing smart meters across the electrical grid. To address these challenges, generating synthetic data has emerged as a promising approach, providing an opportunity to enhance energy efficiency, demand flexibility, and power grid operation. Therefore, this study proposes a nonlinear model of independent component estimation (NICE) with convolutional layers to produce realistic load profiles. This research aims to evaluate the potential of deep generative models (DGMs) through the characterization and quantification of electricity consumption profiles obtained from an actual smart grid on a university campus. The Kullback–Leibler divergence is used to evaluate the performance of the proposed model. Simulation results show that the proposed model can accurately capture the spatiotemporal correlation of actual samples, leading to synthetic load profiles that closely resemble actual profiles. The performance of the proposed NICE model is compared with a NICE model with dense layers, as well as with Generative Adversarial Networks (GAN) with dense layers, and GAN with convolutional layers (cGAN), all methods previously used in the literature to generate synthetic load profiles. It was observed that the proposed NICE model with convolutional layers leads to better results. This model produces more significant similarity between the probability distributions of actual and synthetic data, in addition to a more extraordinary ability to reproduce more realistic load variability curves.