Pattern Blending through Label Interpolation for Smart Meter Data Augmentation with Conditional Generative Adversarial Network

Many applications of smart grids require real-world measurement from smart meter as a foundation, which is however often missing. Unlike some bottom-up methods that model all household devices and simulate energy consumption behaviors, data-drive methods such as generative adversarial networks (GAN) can learn a mapping from a latent space to a sample space, thereby generating a large amount of synthetic data to augment the original dataset. To control the generation process, class labels are provided as an additional input enabling the conditional GAN (cGAN) to generate data of specified classes as required. Instead of focusing only on data from individual classes, this paper employs label interpolation for pattern blending of smart meter data from different classes. The following two questions are discussed: 1) What look like the synthetic load profiles after pattern bending; 2) How and which aspects of the synthetic load profiles change during the blending process. Comprehensive statistical tests prove that label interpolation not only enables a smooth transition between classes but also allows merging characteristics of load profiles across multiple features.