Conditional-TimeGAN for Realistic and High-Quality Appliance Trajectories Generation and Data Augmentation in Nonintrusive Load Monitoring

Non-intrusive load monitoring (NILM) strives to achieve real-time monitoring of individual appliance energy consumption and usage by leveraging aggregate power readings. Most of the existing NILM models are based on machine learning. To improve the generalization and accuracy of NILM models, it is crucial to expand the dataset. However, collecting a large amount of power data is a challenging and common task. To address this, we propose a novel model called Conditional-Time-series Generative Adversarial Networks (C-TimeGAN), which extends upon TimeGAN by incorporating constraint conditions to generate high-quality and realistic appliance trajectories. This model preserves the benefits of unsupervised GANs’ flexibility and supervised training’s stepwise control in TimeGAN while addressing challenges in appliance data generation, like rapid power changes and transients. Additionally, we utilize a one-class support vector machine (OCSVM) for post-processing to further enhance data quality by detecting anomalies and removing outliers. Experiments show the synthetic data from our C-TimeGAN+ model (C-TimeGAN integrated with OCSVM) outperforms the baseline in terms of both quality and quantity, exhibiting higher fidelity, diversity, and novelty. Furthermore, C-TimeGAN+ serves as a practical data augmentation tool, enhancing disaggregation accuracy and generalization of other NILM models. We also discuss the optimal proportion of synthetic data for augmentation, improving the practical applicability of the data.