An academic team has published a fully local AI pipeline that turns noisy appliance energy data into readable maintenance tickets.
The system monitors seven office appliances using a hybrid forecasting model that pairs Singular Spectrum Analysis with Long Short-Term Memory networks. When consumption looks off, a per-appliance LSTM Variational Autoencoder flags the episode. From there, a three-stage LangChain pipeline takes over: a Context Agent pulls relevant background sources, a Diagnosis Agent structures the evidence into JSON, and a Report Agent writes a human-readable summary. Operators can push feedback back into a memory layer, so the system learns from corrections over time.
The practical angle is the retrieval strategy. Instead of dumping every available source into the prompt, the Context Agent starts with three core references and conditionally adds up to three more depending on the event — capping reasoning at eight steps. That dynamic approach matched the performance of full static retrieval while cutting average context roughly in half. The best-performing backend scored 90.4 out of 100 across a 16-scenario benchmark; a fully local 7B-parameter model passed all 16 scenarios, which matters for facilities that cannot or will not route data to a cloud API.
Building automation has promised smart energy management for decades, and most of it has delivered dashboards that non-experts quietly ignore. Routing anomaly explanations through an LLM is not a new idea — but the combination of local inference, operator feedback loops, and capped retrieval at least addresses the three most common reasons these systems fail in practice: privacy concerns, alert fatigue, and explanations written for engineers rather than facility managers.