Domain Adaptation Using Adversarial Neural Network with Correlation Alignment Loss for Household Appliance Classification

Energy monitoring and appliance identification are critical for addressing energy crisis and environmental pollution. However, variations in appliance load profiles across different households and locations pose significant challenges to accurate identification. Moreover, existing models often assume access to sufficient labeled training data, which can be costly or impractical to obtain. To overcome these challenges and enhance cross-domain generalization, we propose a vision transformer-based unsupervised adversarial appliance identification model that incorporates domain adaptation with CORrelation ALignment (CORAL) loss. A pretrained Vision Transformer (ViT) is used as a feature extractor to jointly learn shared representations for both the label predictor and domain classifier, enabling the model to focus on domain-specific characteristics and mitigate domain shift. Low-resolution daily power consumption signals are segmented into 15-minute resolution time-series windows and transformed into Gramian Angular Difference Field (GADF) images to enhance feature extraction. The model is then adversarially trained on source and target domains using a negative-log likelihood loss augmented with correlation alignment to improve stability and feature alignment. We evaluated the model under three different scenarios: (i) source and target domains from different geographic regions, (ii) source and target domains from the same region but different datasets (utilities), and (iii) source and target domains from the same dataset but different houses. Experiments on three public datasets demonstrate substantial improvements, achieving macro F1 score ranging from 12.66% to 70.82% in scenario (i), 2.87% to 84.62% in scenario (ii), and 7.77% to 362.56% in scenario (iii). An ablation study further reveals that incorporating CORAL loss can achieve up to a 75.34% improvement in macro F1 score compared to domain adaptation without it. Additionally, the proposed model consistently outperforms direct time-series models that uses 1D raw signals. It also demonstrates superior scalability, robustness under various input conditions, and efficient computational complexity with faster training and inference times.