Abstract
The ADRENALIN 2024 Load Disaggregation Challenge: Energy Use in Buildings aimed to advance non-intrusive load monitoring (NILM) by developing unsupervised machine learning models capable of separating temperature-dependent heating, ventilation, and air conditioning (HVAC) loads from total building energy consumption. Unlike traditional NILM approaches that rely on labeled training data, this competition required participants to design scalable and generalizable methods without access to appliance-level ground truth data. The top three solutions demonstrated three distinct methodologies: an adjusted Seasonal-Trend decomposition using LOESS (STL) model leveraging reference weeks for seasonal adjustments, a Gaussian Mixture Model (GMM) clustering-based approach, and a base load decomposition technique distinguishing operational and HVAC-related loads. The highest-performing model achieved an average Normalized Mean Absolute Error (NMAE¯) of 0.235, successfully capturing seasonal HVAC trends while maintaining scalability and robustness. The competition results highlighted key challenges in automating HVAC disaggregation, improving model generalization across diverse building types, and integrating additional contextual data sources. Future research should focus on self-supervised learning, multi-modal NILM techniques, federated learning for privacy-preserving disaggregation, and computational efficiency for real-time NILM applications. These advancements will be instrumental in enabling energy-efficient building management, demand response strategies, and smart grid integration.