eHabitat : Training Energy Home Assistants to Reduce their Energy Habitat

Khalid Oublal

Keywords: Non-Intrusive Load Monitoring, NILM, Smart Home Simulation, Energy Disaggregation, Synthetic Data, Appliance-Level Energy Modeling, Benchmarking, Deep Learning, Hierarchical Models

TL;DR: eHabitat NILM is a high-performance simulation platform with realistic 3D home datasets and benchmarks, enabling fast evaluation of NILM algorithms.

Abstract: We introduce eHabitat NILM (eHabitat), a simulation and benchmarking platform for evaluating non-intrusive load monitoring (NILM) algorithms in realistic residential environments. Our contributions span the full stack of smart home energy research: data, simulation, and benchmark tasks. Specifically, we present (i) ApplianceCAD, an annotated, reconfigurable 3D dataset of homes with appliances and their states (on/off, standby, varying power levels) to match real-world energy usage patterns; (ii) eHabitat Simulator, a high-performance physics- and usage-pattern–aware simulator capable of generating synthetic electricity consumption traces at scale, supporting over 25,000 simulated appliance-hours per second, enabling hundreds of times faster experimentation than prior NILM simulators; and (iii) Home Energy Benchmark (HEB), a suite of energy disaggregation tasks covering diverse households and appliance types, designed to test NILM models on generalization across unseen appliances, usage patterns, and home layouts. These contributions enable systematic comparison of deep learning–based NILM models versus classical signal-processing approaches in long-horizon monitoring scenarios. We find that flat neural architectures struggle with generalization across households, hierarchical models with independently trained appliance modules can suffer from handoff errors between appliances, and classical NILM pipelines are less robust than learned models under realistic variability in appliance usage.

Submission Number: 7