Non-Intrusive Load Monitoring (NILM): A Systematic Review of Observability Regimes, Methods, and Evaluation Paradigms

This paper presents a systematic review of Non-Intrusive Load Monitoring (NILM), analyzing its scientific, methodological, and infrastructural evolution over more than three decades. Rather than being a taxonomic survey, the review develops a system-level synthesis in which NILM is conceptualized as an observability-conditioned cyber-physical inference problem. In this formulation, sensing regimes, physical observability ceilings, abstraction mechanisms, and deployment contexts jointly delimit the admissible inference space and condition achievable performance. A corpus of 2,908 publications indexed in the Web of Science (WoS) and Scopus was analyzed in accordance with the Preferred Reporting Items for Systematic reviews and Meta-Analyses (PRISMA) protocol and interpreted through an integrated bibliometric and qualitative analytical framework structured under the Theory-Context-Characteristics-Methodology (TCCM) perspective. The literature is organized into methodological regimes including deterministic low-frequency formulations, probabilistic and optimization-based models, and abstraction-driven deep learning architectures. The analysis indicates that reported performance gains are primarily conditioned by sensing infrastructures and observability structures rather than by algorithmic refinement in isolation. Heterogeneity in sensing configurations and evaluation protocols limits cross-study comparability and deployment transferability. A structural gap is identified between abstraction-driven performance gains and the progressive attenuation of explicit physical interpretability, with implications for validation transparency, physical coherence, and operational robustness. Interpretability is therefore examined as an evaluation-layer property emerging from the interaction between observability regimes, abstraction depth, and deployment constraints. The review concludes that sustainable progress in NILM requires regime-aware design and evaluation paradigms that integrate performance, generalization, physical coherence, and interpretability within deployment-constrained operational environments.