Capability and Explanation Boundaries of Large Language Models for Non-Intrusive Load Monitoring

Non-intrusive load monitoring (NILM) aims to disaggregate total electricity consumption into individual appliance usage. Although deep learning has advanced NILM, it remains limited by its dependence on labeled data, restricted generalization, and lack of explainability. This paper introduces a prompt-based NILM framework that uses large language models (LLMs) with in-context learning. We integrate appliance features, contextual history, and one-shot examples into a unified prompting strategy and evaluate it on the REDD and UK-DALE datasets. Prompt-only LLMs deliver basic NILM capability but lag behind specialized deep learning models in accuracy, while achieving zero-shot generalization across houses and regions by adapting the injected appliance features. To characterize the unique value of LLMs for NILM, we propose a three-layer faithfulness framework that evaluates LLM-generated explanations along factual validity, internal coherence, and external correctness. The framework reveals a traceable-but-misaligned pattern in which LLMs often cite numerical values supported by the input but fail to align the cited events with true appliance activity. Rather than replacing specialized models, LLMs are better positioned as a semantic interface for energy disaggregation. We release our framework to support follow-up research.