Datasets:

shahoismael
/

crossclimatepv-protocol

timestamp timestamp[ms]date 2010-07-01 00:00:00 2013-06-30 23:30:00	site_id stringclasses 300 values	split stringclasses 3 values	is_rare_event bool 2 classes
2010-07-01T00:00:00	ID_1	train	false
2010-07-01T00:30:00	ID_1	train	false
2010-07-01T01:00:00	ID_1	train	true
2010-07-01T01:30:00	ID_1	train	false
2010-07-01T02:00:00	ID_1	train	false
2010-07-01T02:30:00	ID_1	train	true
2010-07-01T03:00:00	ID_1	train	false
2010-07-01T03:30:00	ID_1	train	false
2010-07-01T04:00:00	ID_1	train	false
2010-07-01T04:30:00	ID_1	train	false
2010-07-01T05:00:00	ID_1	train	true
2010-07-01T05:30:00	ID_1	train	false
2010-07-01T06:00:00	ID_1	train	false
2010-07-01T06:30:00	ID_1	train	false
2010-07-01T07:00:00	ID_1	train	false
2010-07-01T07:30:00	ID_1	train	false
2010-07-01T08:00:00	ID_1	train	true
2010-07-01T08:30:00	ID_1	train	false
2010-07-01T09:00:00	ID_1	train	false
2010-07-01T09:30:00	ID_1	train	true
2010-07-01T10:00:00	ID_1	train	false
2010-07-01T10:30:00	ID_1	train	false
2010-07-01T11:00:00	ID_1	train	false
2010-07-01T11:30:00	ID_1	train	false
2010-07-01T12:00:00	ID_1	train	false
2010-07-01T12:30:00	ID_1	train	false
2010-07-01T13:00:00	ID_1	train	false
2010-07-01T13:30:00	ID_1	train	true
2010-07-01T14:00:00	ID_1	train	false
2010-07-01T14:30:00	ID_1	train	false
2010-07-01T15:00:00	ID_1	train	false
2010-07-01T15:30:00	ID_1	train	true
2010-07-01T16:00:00	ID_1	train	true
2010-07-01T16:30:00	ID_1	train	false
2010-07-01T17:00:00	ID_1	train	false
2010-07-01T17:30:00	ID_1	train	true
2010-07-01T18:00:00	ID_1	train	true
2010-07-01T18:30:00	ID_1	train	false
2010-07-01T19:00:00	ID_1	train	false
2010-07-01T19:30:00	ID_1	train	false
2010-07-01T20:00:00	ID_1	train	false
2010-07-01T20:30:00	ID_1	train	false
2010-07-01T21:00:00	ID_1	train	false
2010-07-01T21:30:00	ID_1	train	true
2010-07-01T22:00:00	ID_1	train	false
2010-07-01T22:30:00	ID_1	train	true
2010-07-01T23:00:00	ID_1	train	true
2010-07-01T23:30:00	ID_1	train	false
2010-07-02T00:00:00	ID_1	train	false
2010-07-02T00:30:00	ID_1	train	false
2010-07-02T01:00:00	ID_1	train	true
2010-07-02T01:30:00	ID_1	train	true
2010-07-02T02:00:00	ID_1	train	false
2010-07-02T02:30:00	ID_1	train	false
2010-07-02T03:00:00	ID_1	train	false
2010-07-02T03:30:00	ID_1	train	true
2010-07-02T04:00:00	ID_1	train	false
2010-07-02T04:30:00	ID_1	train	false
2010-07-02T05:00:00	ID_1	train	false
2010-07-02T05:30:00	ID_1	train	true
2010-07-02T06:00:00	ID_1	train	false
2010-07-02T06:30:00	ID_1	train	false
2010-07-02T07:00:00	ID_1	train	false
2010-07-02T07:30:00	ID_1	train	false
2010-07-02T08:00:00	ID_1	train	true
2010-07-02T08:30:00	ID_1	train	false
2010-07-02T09:00:00	ID_1	train	true
2010-07-02T09:30:00	ID_1	train	false
2010-07-02T10:00:00	ID_1	train	false
2010-07-02T10:30:00	ID_1	train	false
2010-07-02T11:00:00	ID_1	train	false
2010-07-02T11:30:00	ID_1	train	false
2010-07-02T12:00:00	ID_1	train	false
2010-07-02T12:30:00	ID_1	train	false
2010-07-02T13:00:00	ID_1	train	true
2010-07-02T13:30:00	ID_1	train	false
2010-07-02T14:00:00	ID_1	train	false
2010-07-02T14:30:00	ID_1	train	false
2010-07-02T15:00:00	ID_1	train	true
2010-07-02T15:30:00	ID_1	train	true
2010-07-02T16:00:00	ID_1	train	false
2010-07-02T16:30:00	ID_1	train	false
2010-07-02T17:00:00	ID_1	train	false
2010-07-02T17:30:00	ID_1	train	false
2010-07-02T18:00:00	ID_1	train	true
2010-07-02T18:30:00	ID_1	train	false
2010-07-02T19:00:00	ID_1	train	true
2010-07-02T19:30:00	ID_1	train	false
2010-07-02T20:00:00	ID_1	train	false
2010-07-02T20:30:00	ID_1	train	false
2010-07-02T21:00:00	ID_1	train	true
2010-07-02T21:30:00	ID_1	train	true
2010-07-02T22:00:00	ID_1	train	false
2010-07-02T22:30:00	ID_1	train	true
2010-07-02T23:00:00	ID_1	train	false
2010-07-02T23:30:00	ID_1	train	false
2010-07-03T00:00:00	ID_1	train	false
2010-07-03T00:30:00	ID_1	train	false
2010-07-03T01:00:00	ID_1	train	true
2010-07-03T01:30:00	ID_1	train	false

End of preview. Expand in Data Studio

CrossClimatePV — protocol labels

Row-level evaluation protocol for CrossClimatePV, a controlled cross-climate benchmark for photovoltaic power forecasting.

This dataset contains no measurements. There is no power output, no irradiance, no temperature, humidity or wind. It contains only labels this project computed: which split each row belongs to, and whether the row is a rare event. That is a deliberate design choice, not an omission — see What this is and is not.

23,312,924 rows across four archives and five Köppen climate zones.

What it is for

CrossClimatePV asks whether a forecasting model that works in one climate still works in another, and — the part no earlier benchmark could answer — why it fails when it does. Reproducing either result requires scoring on exactly the rows the paper scored. Per-site split boundaries get you close; this dataset gets you exact.

Join on (site_id, timestamp) against your own copy of the source archive and you land on the identical evaluation rows, with the identical rare-event labels.

Config	Köppen	Sites	Rows	Rare events
DKASC	BWh	1	423,510	22,966 (5.42%)
HKUST	Cwa	60	4,711,971	164,364 (3.49%)
Ausgrid	Cfa	300	15,778,512	2,529,808 (16.03%)
PVDAQ	BSk, BWh, Cfa, Dfb	8	2,398,931	128,997 (5.38%)
Total	5 distinct classes	369	23,312,924	2,846,135

Those percentages are over the whole record. The paper reports prevalence on the test split of the scored sites, which is what its rare-event results are measured on: 9.25% DKASC, 5.73% HKUST, 16.28% Ausgrid, 5.00% PVDAQ.

345 of the 369 sites are scored. HKUST contributes 37 of 60 and PVDAQ 7 of 8; the rest lack sufficient usable record inside the test window.

Schema

Column	Type	Meaning
`timestamp`	`timestamp[s]`	Start of the interval, in the archive's own local convention
`site_id`	`string`	Anonymous site identifier, matching `protocol/` in the GitHub repo
`split`	`string`	`train`, `val` or `test`
`is_rare_event`	`bool`	Rare event under the paper's threshold setting

How splits were made

Chronological 70/15/15, no shuffling. PVDAQ uses a per-site split because its systems were commissioned across 2007–2023 and a single global window would have excluded most of them; this asymmetry is disclosed in the paper.

How rare events were labelled

Two proxies, combined: an inverter-clipping rule (output pinned near the 99.5th percentile for consecutive intervals) and a cloud-transient rule (output dropping by at least half within one interval). The paper's setting is P99.5, 1% tolerance, 3 steps, 50% drop, chosen from a 405-setting sensitivity sweep.

Worth knowing before you use these labels: the cloud-transient rule carries the entire effect. It is positive in 540 of 540 threshold settings tested. The clipping rule runs the other way — models are better on clipped rows, because a flat ceiling at rated output is trivially predictable. The combined rule survives only because cloud transients dominate it numerically, clipping affecting under 1% of rows at every setting tested.

Usage

from datasets import load_dataset

ds = load_dataset("shahoismael/crossclimatepv-protocol", "PVDAQ", split="train")

Or with pandas, which is usually what you want for a join:

import pandas as pd

labels = pd.read_parquet(
    "hf://datasets/shahoismael/crossclimatepv-protocol/data/Ausgrid.parquet"
)
merged = your_ausgrid_frame.merge(labels, on=["site_id", "timestamp"], how="inner")
test = merged[merged.split == "test"]

What this is and is not

Not included, on purpose. The four source archives are not redistributed here or anywhere else in this project. Each remains under the terms of its original provider and must be obtained from source:

DKASC — Desert Knowledge Australia Solar Centre, Alice Springs. https://dkasolarcentre.com.au/download
HKUST — rooftop PV dataset published in Scientific Data, via Dryad.
Ausgrid — Solar Home Electricity Data, 300 de-identified residential customers. https://www.ausgrid.com.au/Industry/Our-Research/Data-to-share Re-identification is prohibited and none was attempted; only anonymous site identifiers appear here.
PVDAQ — NREL Photovoltaic Data Acquisition, via the Open Energy Data Initiative. https://openei.org/wiki/PVDAQ

Reproducing the corpus. The harmonization code in the GitHub repository rebuilds the 23,312,924-row corpus exactly from those four sources. If your harmonization produces a different row count, it diverged somewhere, and these labels will not join cleanly.

Licence

CC BY 4.0, to the extent this project holds rights in these derived labels. That grant does not extend to the underlying source data. Where a source's terms conflict, the source's terms prevail. See LICENSE-DATA.

Citation

@software{crossclimatepv_2026,
  author  = {Ismael Hassen, Shaho},
  title   = {CrossClimatePV},
  year    = {2026},
  version = {1.1.0},
  doi     = {10.5281/zenodo.21918702},
  url     = {https://github.com/shahoismael/crossclimatepv}
}

The paper — CrossClimatePV: A Controlled Multi-Climate Benchmark Showing Climate Outweighs Every Dataset Artifact Tested in Photovoltaic Power Forecasting — is under review. This card will be updated when it appears.