README.md

---
license: mit
task_categories:
  - time-series-forecasting
language:
  - en
tags:
  - nilm
  - energy-disaggregation
  - uk-dale
  - high-frequency
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: train/*
      - split: val
        path: val/*
      - split: benchmark
        path: benchmark/*
---

# NILMbench processed UK-DALE splits

Pre-processed 16 kHz voltage/current frames and per-category active-power
labels from the UK-DALE 2015 release, packaged for the NILMbench benchmark
(House 1 → House 2 cross-household evaluation).

## Layout

```
train/      10,000 sparse class-balanced 6-second frames from House 1
val/         1,000 sparse class-balanced 6-second frames from House 1
benchmark/   2,000 sparse class-balanced 6-second frames from House 2
```

Each split contains:

| File                     | Shape          | Description                                |
| ------------------------ | -------------- | ------------------------------------------ |
| `x_vi_6s.npy`            | `(N, 2, 96000)` float16 | 16 kHz V/I waveform per frame (FLAC-normalised, range `[-1, 1]`) |
| `labels_and_index.npz`   | dict            | per-category power label, on/off label, aggregate context, timestamp, source window id |

`labels_and_index.npz` contains:

* `y_power`       `(N, 7)` float32 — active power in watts per scored category
* `y_state`       `(N, 7)` bool — on/off label per category
* `x_agg`         `(N, 11)` float32 — aggregate-power context (±5 frames, ±30 s)
* `timestamp`     `(N,)` int64 — Unix seconds of frame centre
* `sample_idx`    `(N,)` int16 — 0..599 index inside the source window
* `window_id`     `(N,)` str  — UK-DALE window identifier
* `class_names`   `(7,)` str  — ordered category names

## Recovering engineering units

The V/I waveforms are stored in FLAC-normalised form (range `[-1, 1]`). To
get volts and amperes, multiply by the UK-DALE House-2 calibration constants:

```python
V_FACTOR = (2 ** 31) * 1.88296904357e-7   # ≈ 404.4
I_FACTOR = (2 ** 31) * 4.77518864497e-8   # ≈ 102.5
```

## Usage

```python
import numpy as np
from huggingface_hub import snapshot_download

root = snapshot_download(repo_id="Pybunny/nilmbench-ukdale", repo_type="dataset")
x = np.load(f"{root}/train/x_vi_6s.npy", mmap_mode="r")
labels = np.load(f"{root}/train/labels_and_index.npz", allow_pickle=True)
print(x.shape, labels["y_power"].shape, labels["class_names"])
```

## Citation

NILMbench paper (2026), and the original UK-DALE dataset by Kelly &
Knottenbelt (2015).

## License

MIT for the processed splits and metadata. The underlying UK-DALE recordings
are subject to their original license (CC-BY 4.0).