PowerZooJax/PowerZooDataset

PowerZoo Dataset

Real-world power-system and data-centre time-series, canonical grid topologies, and JSON manifests linking the two — packaged for reinforcement-learning and forecasting research with the PowerZoo / PowerZooJax benchmark code (released separately).

Submitted to the NeurIPS 2026 Datasets & Benchmarks Track. Author information is omitted during double-blind review.

---

1. What's inside

1. Eleven parquet time-series files ingested from public regulator and cloud-provider releases (GB / AU electricity load, generation by fuel, day-ahead forecasts, market mid-prices; Alibaba / Azure / Google data-centre utilisation).
2. Fourteen electrical-network case files (Python classes with bus / branch / generator / load tables) covering transmission systems from 5 to 2383 buses and distribution systems from 33 to 533 buses.
3. Eleven JSON manifests that map each parquet's raw columns to a shared canonical schema (e.g. OPERATIONAL_DEMAND → load.actual_mw), so traces from different sources compose cleanly in one experiment.

---

2. Repository layout

PowerZooDataset/
├── README.md                  # this file
├── parquet/                   # harmonised time-series traces
│   ├── *.parquet              # primary data files
│   └── *.json                 # per-file provenance metadata (rows, dtypes, source URL, generation timestamp)
├── manifests/                 # loader manifests (column maps, derivations, normalisation)
│   └── <dataset_id>.json
└── powergrid_case/            # electrical-network case definitions (Python)
    ├── CaseBase.py            # ClearCase base class + ext. DataFrame
    ├── transmission/          # HV / sub-transmission test systems
    │   ├── Case5.py, Case14.py, Case29GB.py, Case118.py,
    │   ├── Case300.py, Case552GB.py, Case1354pegase.py, Case2383wp.py
    │   └── __init__.py
    └── distribution/          # MV distribution test feeders
        ├── Case33bw.py, Case118zh.py, Case123.py, Case141.py,
        ├── Case533mt_hi.py, Case533mt_lo.py
        └── __init__.py

2.1 Parquet traces

File	Domain	Resolution	Rows	Columns	Bytes
AEMO_5min_Demand_2025_2026.parquet	AU NEM demand (5 regions)	5 min	737,400	5	6.5 MB
AEMO_Forecast_vs_Actual_2025.parquet	AU NEM probabilistic forecast vs. actual	30 min	89,145	10	1.6 MB
Ausgrid_Zone_Substation_FY25_imputed_15min.parquet	NSW zone substations (175 sites)	15 min	6,095,040	4	60 MB
GB_NESO_Demand_2009_2025_30min.parquet	GB NESO historical demand	30 min	285,454	22	4.7 MB
GB_Forecast_Actual_Demand_2023_2025_30min.parquet	GB day-ahead forecast vs. actual	30 min	48,283	3	0.8 MB
GB_Gen_by_Type_2016_2025_30min.parquet	GB generation by fuel type	30 min	180,048	13	6.2 MB
MID_GB_30min_aligned_to_gen.parquet	GB APX/N2EX mid prices & volumes	30 min	48,283	6	0.8 MB
alibaba_dc_2018_300s.parquet	Alibaba production cluster (CPU/mem/net/disk)	5 min	2,243	6	0.1 MB
alibaba_gpu_2020_300s.parquet	Alibaba GPU cluster (GPU/CPU util)	5 min	415	3	<0.1 MB
azure_dc_v2_300s.parquet	Azure VM trace v2 (CPU, assigned memory)	5 min	8,640	3	0.2 MB
google_dc_2019_300s.parquet	Google Borg 2019 (CPU/mem/CPI)	5 min	8,064	5	0.3 MB

Each *.parquet ships with a sibling *.json capturing source URL, source organisation, generation timestamp, exact column dtypes, region/category enumerations and (where applicable) timezone conventions.

2.2 Manifests

Manifests in manifests/ are the contract used by PowerZoo / PowerZooJax loaders. A manifest declares:

• parquet_file and the matching metadata_json
• column_map: rename rules from raw column → canonical schema (e.g. OPERATIONAL_DEMAND → load.actual_mw)
• index_map: which raw columns serve as datetime / region / issue_time / target_time
• derived: closed-form derivations from raw columns (e.g. wind.available_mw = "Wind Offshore + Wind Onshore")
• normalize: per-channel scaling factors applied at load time
• time_mode: calendar (absolute UTC) or profile (cyclical, anchored to data_epoch)
• region_values, date_range, source_url, source_organization

The 11 manifests cover every parquet file shipped here. Note that gb_neso_demand uses a two-column index (SETTLEMENT_DATE + SETTLEMENT_PERIOD 1–50) instead of a single datetime column — the manifest's datetime_recipe field documents the exact reconstruction (Europe/London tz-localise then convert to UTC; SP=49–50 absorb the autumn clock-change repeat).

2.3 Power-grid case files

powergrid_case/ contains a unified Python representation for both transmission and distribution test systems. Every case subclasses ClearCase and exposes four pandas.DataFrame tables in MATPOWER-compatible units (MW, MVAr, p.u.):

• nodes — id, type (1=PQ / 2=PV / 3=Ref), Pd, Qd, x, y
• units — id, bus_id, mc_a, mc_b, mc_c, p_max, p_min (quadratic cost + capacity)
• lines — id, from, to, x, floor, cap (reactance + thermal limits)
• loads — id, bus_id, mc_a, mc_b, mc_c, d_max, d_min (price-responsive demand)

Each file declares BUS_COUNT, VOLTAGE_LEVEL, SOURCE and DESCRIPTION as class-level metadata. The values below are reproduced verbatim from those declarations (so the file itself is the source of truth):

File	Voltage	Buses	SOURCE	DESCRIPTION
transmission/Case5.py	HV	5	MATPOWER	IEEE 5-bus test system
transmission/Case14.py	HV	14	MATPOWER	IEEE 14-bus test system
transmission/Case29GB.py	HV	29	custom	GB reduced 29-bus transmission network
transmission/Case118.py	HV	118	MATPOWER	IEEE 118-bus test system
transmission/Case300.py	HV	300	MATPOWER	IEEE 300-bus test system
transmission/Case552GB.py	HV	552	GB	Great Britain 552-bus transmission (distinct from 29-bus Case29GB)
transmission/Case1354pegase.py	HV	1354	MATPOWER	European PEGASE 1354-bus system
transmission/Case2383wp.py	HV	2383	MATPOWER	Polish 2383-bus winter peak system
distribution/Case33bw.py	MV	33	MATPOWER	IEEE 33-bus Baran & Wu radial distribution
distribution/Case118zh.py	MV	118	MATPOWER	118-bus Zhang distribution system
distribution/Case123.py	MV	123	MATPOWER	IEEE 123-bus three-phase distribution
distribution/Case141.py	MV	141	MATPOWER	141-bus Caracas distribution system (Khodr et al., EPSR 2008)
distribution/Case533mt_hi.py	MV	533	MATPOWER	533-bus Swedish distribution (high load)
distribution/Case533mt_lo.py	MV	533	MATPOWER	533-bus Swedish distribution (low load)

---

3. Loading

3.1 Direct parquet load (no extra dependency on PowerZoo)

import pandas as pd
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="PowerZooJax/PowerZooDataset",
    repo_type="dataset",
    filename="parquet/AEMO_5min_Demand_2025_2026.parquet",
)
df = pd.read_parquet(path)
print(df.head())

3.2 Via datasets (per-config)

from datasets import load_dataset

ds = load_dataset(
    "PowerZooJax/PowerZooDataset",
    name="aemo_5min_demand",
    split="train",
)

3.3 Via the PowerZooJax DataLoader

If you have the (separately released) benchmark package installed, instantiate DataLoader with this repository's parquet/ and manifests/ directories:

from powerzoojax.data import DataLoader

loader = DataLoader(
    data_dir="/path/to/PowerZooDataset/parquet",
    manifest_dir="/path/to/PowerZooDataset/manifests",
)
print(loader.list_available_datasets())
df = loader.load_actual_series("aemo_5min_demand")

3.4 Power-grid cases

from powerzoo.case.distribution.Case141 import Case141
case = Case141()
case.check()
case.get_node_ptdf()

---

4. Schema conventions

• Time stamps. Calendar-mode parquet files store timestamps as datetime64[ns, UTC], with the exception of GB_NESO_Demand_2009_2025_30min.parquet, which uses two columns (SETTLEMENT_DATE + SETTLEMENT_PERIOD) — see the datetime_recipe in manifests/gb_neso_demand.json. The Ausgrid metadata explicitly declares timezone_local = "Australia/Sydney" and timezone_stored = "UTC"; for the other calendar sources, only the stored UTC representation is documented in metadata.
• Profile-mode traces. The data-centre traces (Alibaba / Azure / Google) are tagged time_mode = "profile" and cyclical = true in their manifests, anchored to a data_epoch (Alibaba DC: 2018-01-06, Alibaba GPU: 2020-07-01, Azure: 2019-01-01, Google: 2019-05-01). They are intended as periodic exogenous signals, not absolute calendar series.
• Imputation. Ausgrid_Zone_Substation_FY25_imputed_15min.parquet contains imputed values (as indicated by its filename suffix and source path); the imputation method is not documented inside the metadata shipped here. Users requiring raw observations should pull from the upstream Ausgrid release.
• Units. Power columns in the parquet traces are in MW. The GB market-mid file uses prices in mid_price_* columns (the upstream Elexon convention is £/MWh; consult the source URL for any unit caveats). Data-centre utilisation columns are 0–100 (percent) before the manifest's normalize factor is applied.

---

5. Source acknowledgements & licensing

This repository redistributes derivative datasets built from publicly accessible upstream releases. Each upstream is governed by its own terms; users are responsible for complying with them. The source_url and source_organization fields in every parquet/*.json and manifests/*.json give the canonical pointer.

Trace	Upstream	Original portal
AEMO 5-min demand & POE forecasts	Australian Energy Market Operator	https://visualisations.aemo.com.au/aemo/nemweb/
Ausgrid zone-substation load	Ausgrid	https://www.ausgrid.com.au/Industry/Our-Research/Data-to-share/Distribution-zone-substation-data
GB demand actual & day-ahead forecast	Elexon (BMRS)	https://data.elexon.co.uk/bmrs/api/v1
GB generation by fuel type	Elexon (BMRS)	https://data.elexon.co.uk/bmrs/api/v1/generation/actual/per-type
GB historical demand 2009–2025	National Energy System Operator (NESO)	https://www.neso.energy/data-portal/historic-demand-data
Alibaba cluster trace 2018	Alibaba Group	https://github.com/alibaba/clusterdata/tree/master/cluster-trace-v2018
Alibaba GPU trace 2020	Alibaba Group	https://github.com/alibaba/clusterdata/tree/master/cluster-trace-gpu-v2020
Azure public dataset v2	Microsoft	https://github.com/Azure/AzurePublicDataset
Google cluster data 2019	Google	https://github.com/google/cluster-data

The packaging artefacts authored for this benchmark (the manifest schema, case-file Python representations, schema harmonisation logic) are intended to be released under a permissive open-source licence at camera-ready time; the underlying parquet data inherits the upstream licence in every case. The canonical copy is the upstream URL above.

All series are aggregated at substation, regional, or national level; no PII.

---

6. Intended use & limitations

Intended use. Reinforcement-learning research, load / generation forecasting, OPF benchmarking, distribution-system control, data-centre power-shaping, demand-response studies.

Limitations.

• Static snapshot at release time; upstream sources continue to publish. Redownload from the URLs in §5 for live data.
• Geographic scope: Great Britain and Australia only.
• The Ausgrid trace is imputed by the upstream publisher; the imputation procedure is not documented here.
• Data-centre traces ship only the columns each manifest maps, at 5-minute resolution; upstream releases offer more fields at finer cadences.
• Grid-case parameter values are consistent with the named source systems but are not byte-identical to any specific upstream release.

Not intended for real-time grid control, retail tariff design, or settlement of real markets.

---

7. File integrity

To regenerate provenance metadata locally:

python -c "
import pyarrow.parquet as pq, glob, os
for p in sorted(glob.glob('parquet/*.parquet')):
    m = pq.read_metadata(p)
    print(f'{os.path.basename(p):60s}  rows={m.num_rows:>10,}  cols={m.num_columns:>3}  bytes={os.path.getsize(p):>10,}')
"

Expected output is reproduced in §2.1.

---

8. Changelog

• v0.1 (NeurIPS 2026 DB submission) — initial public release: 11 harmonised parquet traces, 11 loader manifests, 14 grid-case files (8 transmission + 6 distribution).