Research map / Time series / Foundation models for time series
SolarCube: An Integrative Benchmark Dataset Harnessing Satellite and In-situ Observations for Large-scale Solar Energy Forecasting
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces SolarCube, an ML-ready benchmark dataset combining GOES-16/Himawari-8 geostationary satellite images, satellite-derived solar radiation, and BSRN/SURFRAD ground measurements across 19 study areas, with baseline benchmarks for area-based and point-based solar forecasting.
Problem
Solar radiation forecasting lacks a large-scale, ML-ready dataset integrating geostationary satellite imagery with fine-grained solar radiation and ground station observations, with consistent evaluation metrics across point- and area-based tasks.
Approach
Integrates geostationary satellite bands, a newly generated 15-minute satellite-derived surface radiation product, and in-situ pyranometer/pyrheliometer measurements over 19 continental study areas, and benchmarks ConvLSTM/video transformers (area) and LSTM/Transformer/Informer (point) with consistent relative metrics and cloud-based easy/hard scenarios.
Key results
- Area-based short-term: Divided-st reaches R2 0.792 and rRMSE 22.32% over all samples, versus persistence R2 0.671 and rRMSE 29.09%.
- Point-based short-term: Transformer reaches R2 0.719 and rRMSE 33.74% over all samples, versus persistence rRMSE 38.20%.
- Point-based long-term (day-ahead): LSTM-a reports R2 0.672 and rRMSE 50.56% over all samples, while persistence has rRMSE 73.08%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Divided-st instead of persistence | Area-based short-term forecasting · Area-based short-term forecasting, All | rRMSE: 29.09 -> 22.32 | Helped |
| Use ConvLSTM instead of persistence | Area-based short-term forecasting · Area-based short-term forecasting, Hard | rRMSE: 26.91 -> 21.94 | Helped |
| Use Divided-st instead of persistence | Area-based short-term forecasting · Area-based short-term forecasting, Easy | rRMSE: 29.96 -> 22.69 | Helped |
| Use a deep learning model instead of persistence | Point-based short-term forecasting · Point-based short-term forecasting, All | rRMSE: 38.20 -> 33.74 | Helped |
| Use Transformer instead of persistence | Point-based short-term forecasting · Point-based short-term forecasting, All | rRMSE: 38.20 -> 33.74 | Helped |
| Use Transformer instead of persistence | Point-based short-term forecasting · Point-based short-term forecasting, Hard | rRMSE: 39.08 -> 34.33 | Helped |
| Use Transformer instead of persistence | Point-based short-term forecasting · Point-based short-term forecasting, Easy | rRMSE: 37.52 -> 33.26 | Helped |
| Use LSTM-a instead of persistence | Point-based long-term forecasting · Point-based long-term forecasting, All | rRMSE: 73.08 -> 50.56 | Helped |
| Use deep learning models instead of persistence | Point-based long-term forecasting · Point-based long-term forecasting, Hard | rRMSE: 100.68 -> 65.31 | Helped |
| Use LSTM-a instead of persistence | Point-based long-term forecasting · Point-based long-term forecasting, Easy | rRMSE: 67.56 -> 47.39 | Helped |
| Use Transformer instead of persistence | Point-based long-term forecasting · Point-based long-term forecasting, Daily | rRMSE: 41.68 -> 32.18 | Helped |
Surprising findings
- Video-swin shows the reverse trend in forecasting skill, improving with horizon up to 1.5 hours before declining.
- Deep learning models show no significant improvement from hard to easy cases in point-based short-term forecasting.
Limitations
- Data covers only the year 2018, limiting temporal generalizability testing.
- Uneven global coverage of ground measurements, with no coverage in Europe.
- Fairness across geographic regions is not addressed in the metrics.
Experimental setup
Datasets: GOES-16 geostationary satellite bands (VIS047, VIS086, IR1330) Himawari-8 geostationary satellite bands satellite-derived physics-based solar radiation (15-min) BSRN ground measurements SURFRAD ground measurements MODIS MCD12 land cover GOES-R and NWC SAF cloud masks ERA5
Benchmarks: SolarCube area-based short-term SolarCube point-based short-term SolarCube point-based long-term
Metrics: R2 rMBD rRMSE forecasting skill (FS)
Baselines: Persistence ConvLSTM Axial attention Video-swin Divided-st LSTM LSTM-a Informer Transformer ERA5
Builds on: ConvLSTM Earthformer Informer Persistence forecasting
Closest papers by meaning
- Improving *day-ahead* Solar Irradiance Time Series Forecasting by Leveraging Spatio-Temporal Context (NeurIPS 2023 · similarity 0.89)
- ml4xcube: Machine Learning Toolkits for Earth System Data Cubes (AAAI 2025 · similarity 0.84)
- Solar Power Generation Forecasting via Multimodal Feature Fusion (Student Abstract) (AAAI 2024 · similarity 0.82)
- Terra: A Multimodal Spatio-Temporal Dataset Spanning the Earth (NeurIPS 2024 · similarity 0.81)
- ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction (NeurIPS 2024 · similarity 0.79)
- DroughtSet: Understanding Drought Through Spatial-Temporal Learning (AAAI 2025 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.