[Paper Review] PowerGAN: Synthesizing Appliance Power Signatures Using Generative Adversarial Networks

Alon Harell, Richard Jones+2 more|arXiv (Cornell University)|Jul 20, 2020

Smart Grid Energy Management39 references4 citations

TL;DR

PowerGAN introduces a conditional, progressively growing 1D Wasserstein GAN to synthesize realistic, diverse appliance power signatures without relying on handcrafted models. It outperforms existing synthetic data methods in both qualitative realism and quantitative metrics like Inception Score (3.81±0.13), FID (43.30), and SWD, demonstrating its effectiveness in generating high-fidelity, truly random appliance power traces for non-intrusive load monitoring (NILM) data augmentation.

ABSTRACT

Non-intrusive load monitoring (NILM) allows users and energy providers to gain insight into home appliance electricity consumption using only the building's smart meter. Most current techniques for NILM are trained using significant amounts of labeled appliances power data. The collection of such data is challenging, making data a major bottleneck in creating well generalizing NILM solutions. To help mitigate the data limitations, we present the first truly synthetic appliance power signature generator. Our solution, PowerGAN, is based on conditional, progressively growing, 1-D Wasserstein generative adversarial network (GAN). Using PowerGAN, we are able to synthesise truly random and realistic appliance power data signatures. We evaluate the samples generated by PowerGAN in a qualitative way as well as numerically by using traditional GAN evaluation methods such as the Inception score.

Motivation & Objective

• To address the critical bottleneck in non-intrusive load monitoring (NILM) caused by limited availability of labeled, real-world appliance power data.
• To develop a data generation method that produces truly random, realistic appliance power signatures without relying on hand-designed physical models.
• To improve NILM generalization by enabling large-scale, diverse synthetic data generation for training and evaluation.
• To evaluate synthetic power traces using established GAN evaluation metrics adapted for time-series power signal data.

Proposed method

• PowerGAN employs a conditional, progressively growing 1D Wasserstein GAN (WGAN) architecture to generate realistic appliance power signatures.
• The generator takes random latent codes and conditional labels (e.g., appliance type) as input to produce time-series power traces.
• The discriminator is trained to distinguish between real power traces (from datasets like ECO and Tracebase) and generated ones, using a 1D ResNet-based classifier for feature extraction.
• Training alternates between updating the generator and discriminator, with the generator minimizing the Wasserstein distance to real data distributions.
• Evaluation uses adapted GAN metrics: Inception Score (IS), Frechet Inception Distance (FID), and Sliced Wasserstein Distance (SWD) on both classifier features and Laplacian pyramid features.
• The method is trained on real power data from ECO and Tracebase to ensure realistic distribution modeling, avoiding reliance on unpublished or synthetic baselines.

Experimental results

Research questions

• RQ1 Can a GAN-based approach generate realistic, diverse appliance power signatures without relying on physical or handcrafted models?
• RQ2 How does PowerGAN compare to existing synthetic data generators (e.g., ANTgen, SynD) in terms of realism and diversity of generated power traces?
• RQ3 To what extent do standard GAN evaluation metrics (IS, FID, SWD) reflect the quality of synthetic appliance power signals?
• RQ4 Can the proposed method generate truly random, novel appliance power signatures not present in training data?

Key findings

• PowerGAN achieves the highest Inception Score (3.81±0.13) among all evaluated generators, indicating superior sample quality and diversity.
• It records the lowest Frechet Inception Distance (43.30), showing the closest distribution match between generated and real power signal features.
• The Sliced Wasserstein Distance using Laplacian features (18±0.088) and classifier features (0.25±0.011) confirms its strong distributional similarity to real data.
• PowerGAN outperforms ANTgen and SynD across all evaluation metrics, demonstrating its superiority in generating realistic and diverse appliance power signatures.
• The use of a 1D ResNet classifier trained on real data enables reliable, unbiased evaluation of synthetic signal quality.