GAN-based Residential Load Data Generation Model Considering Users’ Privacy

Abstract

Residential load is one of the main parts of the power system load and a critical part of the power system planning and operation management, supporting meaningful research on residential load characteristics, load aggregated control, and demand-side response. However, obtaining the residential load data in recent years has become more challenging due to personal privacy concerns. Using existing data to generate new data becomes one of the main ways to obtain load data nowadays. This paper proposes a residential load data generation model based on Wasserstein Conditional Generative Adversarial Network with gradient penalty, which contains independently trained generator and discriminator. The model introduces Wasserstein distance and gradient penalty to the traditional Generative Adversarial Network, which can generate high-quality load data with relatively small training data volume, and improves the difficult and unstable training of the traditional Generative Adversarial Network. The analysis of cases from an Irish smart meter trial dataset shows that the model can learn users' real load data features without involving users’ privacy.