Non-intrusive load monitoring for smart residential buildings: An AI-based framework with cross-country generalization

Accurate appliance-level electricity monitoring is essential for smart home energy management, demand response, occupant feedback, and building-level efficiency optimization. However, many existing non-intrusive load monitoring (NILM) approaches are evaluated within single datasets and controlled environments, limiting their demonstrated generalizability across different households, countries, and measurement conditions. This study presents an AI-based NILM framework designed for real-time smart home applications and edge-integrated deployment. The proposed framework operates on aggregate household electricity signals and provides both appliance-level power estimation and operational state detection within a unified system. Unlike prior works that focus primarily on algorithmic performance within isolated datasets, this study performs extensive cross-dataset and cross-country evaluations using four publicly available datasets from different regions (REDD - USA, UK-DALE - UK, REFIT - UK, and ENERTALK - South Korea). To the best of our knowledge, this is among the first works to systematically examine the generalizability of a deployable NILM framework across geographically diverse residential environments. Results demonstrate competitive performance across datasets in a zero-shot setting, while highlighting the challenges of cross-domain generalization under varying sampling rates, appliance compositions, and user behavior patterns. In addition to accuracy, the framework is evaluated on an edge gateway platform, confirming real-time inference feasibility on a Raspberry Pi 4-class device for practical residential deployment. These findings support the practical integration of NILM into smart residential buildings, enabling scalable and energy-efficient deployment across heterogeneous global settings.