ViT4LPA: A Vision Transformer for advanced smart meter Load Profile Analysis

This paper introduces ViT4LPA, a novel application of Vision Transformer (ViT) technology tailored for advanced Load Profile Analysis (LPA). By converting smart meter load profiles, alongside temperature and irradiance data, into color-coded images, we harness ViT’s potent image processing capabilities for LPA. A large-scale ViT model undergoes pre-training via masked image modeling tasks, utilizing an extensive smart meter dataset from over 4,000 residential households spanning two years. This self-supervised learning approach allows the ViT network to reconstruct original images from masked inputs based on selected masking strategies, enabling the model to discern relationships between load patterns and temperature sensitivity in the load image. This process yields richly informative load embeddings for each customer. Following a thorough pre-training phase, we apply the ViT model to downstream tasks, with a focus on Heating, Ventilation, and Air Conditioning (HVAC) load disaggregation and electric vehicle (EV) and photovoltaic (PV) load identification. Here, the model excels in both tasks outperforming several leading benchmarks demonstrating the ViT4LPA model’s superior performance in capturing load characteristics and reducing errors in load estimation and load identification. Furthermore, an extensive analysis of the ViT4LPA network components, including its positional embeddings and attention mechanisms, offers deep insights into the model’s operational dynamics and its strategic approach to LPA.