Monicahhhhh06/Solar-UniBench-v1.0

Dataset Card for Solar-UniBench v1.0

Dataset Summary

Solar-UniBench v1.0 is an extensible, analysis-ready multimodal solar forecasting data product that harmonizes public observations from high solar penetration regions (California, Europe, and southeastern Australia) through an agentic physics-guided curation pipeline into a shared, quality-controlled record for reproducible forecasting evaluation.

The release covers 11 sites and 91 station-years, with source periods spanning 2008 to 2022 at native temporal resolutions of 1 min, 30 min, and 1 h. It contains tabular solar/meteorological measurements and more than 1.36 million sky camera and satellite images.

• Paper: Solar-UniBench v1.0: Harmonizing heterogeneous public solar data into a spatiotemporally aligned and quality-aware data product
• Repository: https://github.com/Maoxuehui/solar-unibench
• Authors: Keyu Qi, Xuehui Mao, Dazhi Yang, Yinghao Chu

---

Dataset Structure & Sub-datasets

Solar-UniBench v1.0 harmonizes four widely used source benchmarks into a unified schema:

Sub-dataset	Location / Sites	Period	Resolution	Target	Modalities
gefcom2014	3 anonymous zones, Southeastern Australia	2012–2014	1 hour	Normalized PV Power (0–1)	Normalized solar power; 12 ECMWF NWP variables
skippd	Stanford University, CA, USA	2017–2019	1 min	Normalized PV Power (0–1)	ERA5 (130 ch) & GFS (135 ch) NWP; $64\times 64$ RGB Fisheye sky images
folsom	Folsom, CA, USA ($38.67^\circ\text{N}, 121.15^\circ\text{W}$)	2014–2016	1 min	GHI ($\text{W/m}^2$)	GHI, DNI, local weather (T, RH); Ground sky images; GOES-15 satellite patches
sunlake	6 BSRN stations (CAB, CNR, IZA, PAL, PAY, TAM)	2008–2022	30 min	GHI ($\text{W/m}^2$)	GHI, DNI, DHI, surface pressure; 11-channel EUMETSAT RSS satellite imagery ($64\times 64$)

---

Curation Pipeline & Features

Solar-UniBench v1.0 provides the following physical data governance enhancements:

1. Geospatial & Timezone Recovery: Missing location coordinates (e.g., in GEFCom2014) and timezone offsets are algorithmically inferred by inverse-matching daytime irradiance curves against theoretical clear-sky trajectories.
2. Physics-Guided Quality Auditing: All timestamps are aligned to a monotonic UTC grid. Range conditioning and nighttime zeroing governed by solar zenith angle ($\cos\theta_z \le 0$) are applied.
3. Clear-Sky Physical Anchoring: Solar irradiance is anchored to Ineichen–Perez clear-sky reference envelopes, calculating the bounded clearness index $k_t \in [0, 1.5]$.
4. Physics-Informed Visual Features: Includes derived Normalized Red-to-Blue Ratio (NRBR) maps ($[\text{R, G, B, NRBR}]$) for cloud opacity detection and dynamic spatial reliability maps based on dense optical flow fields.

---

Data Schema

1. Tabular Schema (Parquet)

• entity_id: Station/zone unique identifier
• timestamp_utc: Monotonic UTC physical clock index (ISO 8601)
• pv_power_normalized: Capacity-normalized generation target $[0, 1]$
• ghi_actual, dni_actual, dhi_actual: Measured ground GHI/DNI/DHI ($\text{W/m}^2$)
• ghi_clearsky, dni_clearsky, dhi_clearsky: Ineichen–Perez model clear-sky references
• clearness_index: Bounded atmospheric index $k_t = \text{clip}\left(\frac{\text{GHI}{\text{obs}}}{\text{GHI}{\text{cs}} + \epsilon}, 0, 1.5\right)$
• cos_zenith, solar_azimuth: Instantaneous solar geometry parameters
• quality_flag: Audit flag (0: Normal, 1: Sensor Anomaly, 2: Spike, 3: Night)
• split: Chronological experiment partition (train: 70%, val: 15%, test: 15%)
• obs_*: Optional local surface meteorological covariates (e.g., Temperature, Relative Humidity, Pressure)
• nwp_*: Co-indexed Numerical Weather Prediction forecast channels

2. Visual Tensors (Zarr) [Optional / Multi-modal]

Raw visual streams are chunked and optimized using Zarr arrays to enable high-throughput batch loading during multimodal model training:

Sub-dataset	Tensor Shape ($T \times H \times W \times C$)	Channel Composition ($C$)	Codec / Compression	Description
folsom	$1,180,000 \times 128 \times 128 \times 3$	$[\text{R}, \text{G}, \text{B}]$	JPEG (quality=90)	Hemispherical ground sky camera images
skippd	$250,598 \times 64 \times 64 \times 3$	$[\text{R}, \text{G}, \text{B}]$	Blosc LZ4 (clevel=5)	Ground fisheye sky camera images
sunlake	$222,383 \times 64 \times 64 \times 11$	11 Multi-spectral channels	Blosc Zstd (clevel=3)	EUMETSAT SEVIRI satellite imagery crops

Note on Feature Pipeline: On-the-fly dynamic preprocessing script concatenates raw RGB channels with the physics-informed NRBR (Normalized Red-to-Blue Ratio) feature map into a standardized 4-channel visual tensor $[\text{R}, \text{G}, \text{B}, \text{NRBR}]$ ($\mathbb{R}^{4 \times H \times W}$).

---

Benchmark Usage & Citation

If you use Solar-UniBench v1.0 in your research, please cite:

@article{qi2026solarunibench,
  title={Solar-UniBench v1.0: Harmonizing heterogeneous public solar data into a spatiotemporally aligned and quality-aware data product},
  author={Qi, Keyu and Mao, Xuehui and Yang, Dazhi and Chu, Yinghao},
  journal={Earth System Science Data (Submitted)},
  year={2026}
}