NILMbench

NILMbench · energy disaggregation

The paper is fixed.
The benchmark keeps moving.

NILMBench2026 reports the BuildSys study. The leaderboard records later, provenance-checked runs. They are related, but they are not the same artifact.

Fixed research artifact · BuildSys 2026

Read the paper

The original study: 16 models, three public datasets, two resolutions, and T1/T2/T3 tests of within-building, cross-building, and cross-dataset transfer.

Models
16
Datasets
3
Tasks
T1–T3
Open paper page PDF Slides

Generated from immutable result bundles

View the leaderboard

New models appear here after their result files pass source, dataset, container, protocol, and seed checks. Smoke and full results stay visibly separate.

Models
—
Aggregates
—
Source runs
—

Reading leaderboard.json…

Open leaderboard Download CSV Protocol audit

Run or extend it

The repository is the benchmark runner.

Use Python 3.11 for development. Real datasets stay on your machine.

git clone https://github.com/nilmtk/nilmbench.git
cd nilmbench
uv venv --python 3.11
uv pip install ".[benchmark]"
nilmbench doctor
Full run instructions Add a model

Where each concern belongs

Four repositories, four jobs.

NILMTKLoad and convert energy data nilmtk-contribImplement and test models NILMbenchRun fixed protocols and publish results NILM MetadataUse one appliance and meter vocabulary