Residential Load Data Generation

Abstract

Listen

Translate article

Due to the technical limitations of metering and privacy concerns of customers, the large-scale and real-time collection of residential load data still remains a big challenge. To address the problem, we use the generative adversarial networks (GANs) to produce synthetic residential loads as an alternative. Different from existing load generation models, the GAN model is based on deep neural networks (DNNs). It includes a generator network that outputs synthetic loads and a discriminator network that differentiates the real or fake loads. Taking advantage of the learning ability of DNNs, we can capture hidden features of the load pattern and describe them accurately. In this chapter, we conduct an investigation of frequently-used GAN variants accounting for their performance at generating residential load. We design the architectures and training methods for different GANs and propose different metrics to evaluate the model performance comprehensively. Case studies demonstrate that the auxiliary classifier GAN (ACGAN) outperforms other models on the real load data from an Irish smart meter trial. It is practical to use the ACGAN to generate synthetic residential loads when in shortage of real data.