| Home Energy Management Systems (HEMS) can improve household energy awareness, efficiency, and flexibility, but many interfaces remain difficult for non-expert users to understand and act upon. Natural-language interaction could lower this barrier, yet cloud-based assistants introduce privacy, cost, and connectivity concerns, while large language models are poorly suited to the resource limits of low-cost edge hardware. This paper presents a privacy-preserving home energy advisory system for Raspberry Pi-class devices that assigns a small language model (SLM) a deliberately bounded role: lightweight non-generative components interpret the query, deterministic functions retrieve and analyze structured energy data, and the SLM renders the grounded result as a concise natural-language answer. This separation preserves auditability, limits hallucination risk, and reduces computational overhead. The system is advisory rather than autonomous: it explains electricity prices, photovoltaic generation, household load status, and suggested operating windows for deferrable appliances without directly controlling devices. We implement a proof-of-concept prototype and define an evaluation centered on routing reliability, runtime feasibility, memory footprint, and the practical value of bounded SLM use in a privacy-preserving edge HEMS setting. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.