GridCL for fine-grained load profiling in smart grids under limited labels.

Abstract

Fine-grained load profiling is important for demand response and energy management in smart grids, yet supervised approaches remain constrained by the scarcity of high-quality labeled datasets. To address this limitation, we propose GridCL, a self-supervised contrastive learning framework for low-label load profiling in smart grids. GridCL forms paired daily-load views using conservative input perturbations-small temporal rolling, multiplicative perturbation, and energy renormalization-and combines them with a temporal convolutional encoder to learn discriminative representations from unlabeled data. Experiments on three anonymized city datasets and one pooled benchmark show that GridCL achieves strong clustering quality on the pooled AllCities benchmark, reaching 0.648±0.115 ARI, 0.719±0.064 NMI, and 0.620±0.046 silhouette, while also attaining a best city-level ARI of 0.804±0.084. Under sparse-label evaluation, GridCL reaches 0.845±0.042 accuracy on the pooled benchmark with only 10% labeled users, and remains stable at 20% and 30% labeled users with accuracies of 0.851±0.033 and 0.849±0.034, respectively. These results indicate that GridCL provides an effective low-label solution for fine-grained load profiling in practical smart-grid settings.