SolarBench: A global solar energy nowcasting benchmark
Authors: Yuhao Nie, Stephen Campbell, Quentin Paletta, Liwenbo Zhang, Tao Jing, Samer Chaaraoui, Jonathan Giezendanner, Andea Scott, Tao Sun, Cong Feng, Max Aragon, Jacques Camier, Adam Jensen, Florian Kotthoff, Yuexing Yang, Yang Ming, Mengying Li, Stefanie Meilinger, Yupeng Wu, Adam Brandt, Sherrie Wang
Organizations: Institute for Data, Systems, and Society, Massachusetts Institute of Technology, United States. · Laboratory for Information and Decision Systems, Massachusetts Institute of Technology, United States. · Department of Electrical Engineering and Computer Science, Massachusetts Institute of Technology, United States. · Φ-lab, ESRIN, European Space Agency, Italy. · Climate Office, ECSAT, European Space Agency, United Kingdom. · Department of Architecture and Built Environment, University of Nottingham, United Kingdom. · Department of Mechanical Engineering, Hong Kong Polytechnic University, Hong Kong. · International Center for Sustainable Development, Hochschule Bonn-Rhein-Sieg University of Applied Sciences, Germany. · Department of Energy Science and Engineering, Stanford University, United States. · Department of Civil and Environmental Engineering, Stanford University, United States. · Department of Mechanical Engineering, The University of Texas at Dallas, United States. · Mines Paris, Universit´e PSL, France. · Independent Researcher, United States. · Department of Civil and Mechanical Engineering, Technical University of Denmark, Denmark. · OFFIS - Institute for Information Technology, Germany. · Department of Mechanical Engineering, Massachusetts Institute of Technology, United States.
As the share of solar power grows, nowcasting weather-driven solar variability becomes critical for reliable energy system operation. State-of-the-art approaches increasingly apply deep learning to sky camera and geostationary satellite observations, but fragmented datasets and inconsistent evaluation make it difficult to determine whether reported improvements generalize across climates, cloud regimes, and photovoltaic (PV) systems. Here we introduce SolarBench, an open global benchmark for image-based solar nowcasting. SolarBench harmonizes more than six million sky and satellite images from 11 diverse sites spanning a decade, together with irradiance or PV output and auxiliary atmospheric data. An accompanying toolbox supports reproducible data access, processing, model development, and evaluation. Using SolarBench, we benchmark representative models and reveal a gap between average forecasting accuracy and the ability to capture rapid solar fluctuations. We further quantify predictability across cloud regimes and demonstrate data-efficient adaptation to new PV systems. SolarBench provides an extensible foundation for fair comparison and methodological innovation in solar nowcasting.
Figures & tables
Figure 1: Conceptual framework of SolarBench . SolarBench provides a unified benchmark infrastructure for solar nowcasting that combines a curated multimodal database and an open machine learning toolbox for standardized data access, processing, model development, and evaluation. By combining a multi-continental harmonized database with reproducible benchmarking pipelines, SolarBench enables fair cross-site evaluation, reproducible research, and scalable development and cross-site adaptation of solar nowcasting systems.
Figure 2: Dataset characteristics . a , Geographic distribution of the 11 SolarBench sites overlaid on long-term mean global downward surface solar radiation from ERA5 for 1980–2020. b , Short-term solar variability across sites, characterized by distributions of absolute 5-min changes in clear sky index ( ∣ΔCSI∣ ). Black diamonds indicate the root mean square of ∣ΔCSI∣ , defined as the variability index, and percentages indicate the fraction of 5-min intervals with ∣ΔCSI∣>0.1 , used here to identify large short-term solar fluctuations. c , Availability and temporal coverage of the three data categories across sites: ground-based observations, remote sensing data, and atmospheric model products, with irradiance or PV output indicated as the forecast target. Labels within the bars indicate the temporal resolution of the corresponding data.
Figure 3: Benchmarking of level and ramp prediction performance across lead times and locations. a , Aggregated forecasting performance. (i) Level prediction performance quantified by forecast skill (%), defined as the improvement in RMSE related to persistence (i.e., persistence has a forecast skill of zero). Solid lines denote the median across locations, and shaded bands denote the interquartile range. (ii) Ramp recall, defined as the fraction of observed ramp events successfully captured within a 30% relative error threshold, aggregated across sites in the same manner. b , Example irradiance forecasts from representative models at a 60 min lead time under relatively stable and highly variable conditions. The stable case highlights the competitiveness of persistence under limited variability, whereas the highly variable case shows that image-based deep learning models achieve better level prediction accuracy but produce smoother forecasts that miss many rapid irradiance transitions, revealing a gap between aggregate accuracy and ramp event detection. Observed and successfully predicted ramp events are highlighted. c , Site-level forecasting performance. (i) Best-performing model at each site and lead time for forecast skill and ramp recall. Each circle represents a site–lead time pair, and marker color denotes the best-performing model. (ii), Irradiance forecasting RMSE of the ViT model across sites and lead times. Locations are sorted by RMSE at the 10 min lead time.
Figure 4: Predictability across sites and cloud regimes . a , Relationship between forecasting error and persistence-based predictability across sites for 10-, 60-, and 120-min lead times. Persistence rRMSE is used as a proxy for forecasting difficulty, with higher values indicating lower predictability. Results are shown for representative models (MLP, OmniVision, and ViT), with Pearson correlation coefficients reported for each lead time. b , Persistence-based predictability across cloud regimes and lead times. Bar plots show persistence rRMSE across sites for Clear, Thin, Thick, and Patchy cloud regimes at 10-, 60-, and 120-min lead times. Solid horizontal lines indicate category median. Dashed lines in the Patchy panels denote the overall persistence error pattern across all cloud regimes, illustrating its close alignment with the cross-location error pattern under Patchy conditions. Representative sky image sequences for each cloud regime are shown above the first row.
Figure 5: Towards scalable deployment of solar forecasting systems. a , Scaling of local forecasting performance with increasing training data across sites with different predictability levels. Deep learning models generally require approximately six months of local observations to achieve break-even performance relative to persistence, while more stable forecast skill typically emerges after ∼ 12 months of training data. Sites with higher predictability generally require more local data before deep learning models outperform persistence. b , Transfer learning from geographically distributed irradiance forecasting datasets to two real-world PV systems: a 30 kW rooftop PV installation at Stanford University, USA, and a 3.275 MW hybrid PV system at the Gatton Solar Research Facility at the University of Queensland, Australia. Pre-training on geographically distributed irradiance data followed by local fine-tuning substantially reduces local PV data requirements while maintaining competitive forecasting performance.
Appendix figures & tables18 assets
Supplementary material from the paper’s appendix.
Appendix
Supplementary Figure 1: Example solar irradiance variability and corresponding cloud cover observations . Left: measured GHI over a representative day. Three time points (i–iii) representing distinct irradiance conditions are highlighted. Right: corresponding ground-based sky images (left column) and geostationary satellite images (right column). The red circle indicates the location of the sky imaging system within each satellite scene. These examples illustrate how cloud observations from ground-based sky cameras and geostationary satellites provide information about short-term variability in surface solar irradiance.
Supplementary Figure 2: Ground-based measurement device setup at the Akwatia site. The measurement system consists of (1) a GeoVision all-sky camera and (2) a Kipp & Zonen RT1 radiometer for GTI measurements. The inset shows a close-up view of the RT1 radiometer. The Kologo site uses the same measurement setup.
Supplementary Figure 3: Ground-based measurement device setup at the Sankt Augustin site. The system includes (1) a GeoVision all-sky camera, (2) a SOLYS2 sun tracker, (3) an SP Lite2 pyranometer for global horizontal irradiance (GHI) measurements, (4) a CMP11 pyranometer for global horizontal irradiance (GHI) measurements, (5) a GHP1 pyrheliometer for direct normal irradiance (DNI) measurements, and (6) a CMP11 pyranometer for diffuse horizontal irradiance (DHI) measurements. Instruments (3)–(6) are mounted on the SOLYS2 sun tracker. Enlarged views of the all-sky camera and the solar radiation measurement system are shown in the lower panels.
Supplementary Figure 4: Ground-based measurement device setup at the Hong Kong site. The system includes (1) a Vivotek FE9380-HV network camera for capturing sky images, (2) a Delta-T SPN1 pyranometer measuring GHI and DHI, from which DNI is calculated, and (3) a Lufft WS500 weather station providing measurements of air temperature, relative humidity, atmospheric pressure, wind speed, and wind direction.
Supplementary Figure 5: Ground-based measurement device setup at the Nottingham site. Panel (a) shows the rooftop site at the University of Nottingham’s Jubilee Campus and the locations of Base A and Base B, enlarged in panels (b) and (c). The system includes (1) a Mobotix Q26B fisheye camera, (2) a Kipp & Zonen RaZON+ sun tracker with a PR1 pyranometer for DHI and a PH1 pyrheliometer for DNI, from which GHI is derived, (3) a CMP10 pyranometer for GHI cross-validation, and (4) a Gill Instruments GMX600 weather station.
Supplementary Figure 6: Relationship between forecasting error and the variability index. rRMSE of the MLP, OmniVision, and ViT models is plotted against the variability index for the 10-, 60-, and 120-min lead times. Each point represents one SolarBench site, and solid lines indicate least-squares linear fits. Pearson correlation coefficients between rRMSE and the variability index are reported for each model. The variability index is computed using a fixed 5-min interval following Stein et al. [ 47 ] for all forecast lead times. The weak and inconsistent correlations across models and lead times indicate that this aggregate variability metric alone does not adequately characterize forecasting difficulty.
Supplementary Figure 7: Distribution of cloud regimes across SolarBench sites. Stacked bar charts showing the fraction of samples assigned to the four cloud regimes (Clear, Thin, Patchy, and Thick) at each site. Percentages denote the proportion of samples in each regime, highlighting substantial differences in cloud regime composition across the benchmark locations.
Supplementary Figure 8: Representative cloud regime examples across SolarBench sites. Sky image sequences illustrating clear, thin cloud, patchy cloud, and thick cloud conditions at Puy de Dôme, Akwatia, Golden, and Nottingham. Each example shows a 15-min historical sequence at 5-min intervals ( t−15 , t−10 , t−5 , and t ), highlighting cross-site differences in the visual appearance of each cloud regime.
Site
Dataset Name (on Hugging Face)
Lat., Lon. ( ∘ )
Köppen–Geiger Climate Class
Data Period
Duration (years)
Akwatia, Ghana
EnerSHelF-Akwatia
6.047, -0.797
Aw: Tropical savanna with dry winters
2022-04-02 – 2022-09-23
0.5
Folsom, USA
UCSD-Folsom
38.642, -121.148
Csa: Hot-summer Mediterranean
2014-01-02 – 2016-12-30
3.0
Gatton, Australia
UQueensland
-27.559, 152.343
Cfa: Humid subtropical
2020-01-23 – 2020-05-06
0.3
Golden, USA
SRRL-BMS
39.742, -105.180
Dfb: Warm-summer humid continental
2021-04-17 – 2024-03-22
2.9
Hong Kong, China
HK-PolyU
22.305, 114.180
Cwa: Dry-winter humid subtropical
2023-02-18 – 2023-12-31
0.9
Kologo, Ghana
EnerSHelF-Kologo
10.737, -1.078
Aw: Tropical savanna with dry winters
2022-04-02 – 2022-10-04
0.5
Appendix
Supplementary Table 1: Site Characteristics.
Site
Sources
Temp. Res. (min)
Size (GB)
Samples
Available Data
Akwatia, Ghana
SolarBench
1
0.6
89,953
Sky image, GTI
Folsom, USA
[ 28 ]
1
5.9
765,140
Sky image, GHI, DNI, DHI, Air temperature, Relative humidity, Pressure, Wind speed, Wind direction, Max wind speed, Precipitation
Gatton, Australia
[ 40 ]
1
0.4
69,347
Sky image, PV output
Golden, USA
[ 26 ]
1
2.9
367,878
Sky image, GHI, DNI, DHI, Air temperature, Relative humidity, Pressure, Wind speed, Wind direction, Max wind speed, Precipitation
Hong Kong, China
SolarBench
1
1.5
223,812
Sky image, GHI, DHI, Air temperature, Relative humidity, Pressure, Wind speed, Wind direction
Kologo, Ghana
SolarBench
1
0.7
92,182
Sky image, GTI
Appendix
Supplementary Table 2: Details of the ground-based dataset for each site.
Supplementary Table 3: Details of geostationary satellite imagery data for each site.
Description
Variable Name
Unit
Total precipitation
tp
mm
10m U wind component
u10
m/s
10m V wind component
v10
m/s
100m V wind component
v100
m/s
100m U wind component
u100
m/s
2m temperature
t2m
K
Appendix
Supplementary Table 4: Processed ERA5 variables.
Description
Variable Name
Unit
Precipitable water for entire atmosphere
precipitable_water_entire_atmosphere
kg m -2
Relative humidity 2m above ground
relative_humidity_2m_above_ground
%
Specific humidity 2m above ground
specific_humidity_2m_above_ground
Mass fraction
Temperature 2m above ground
temperature_2m_above_ground
∘ C
U component of wind 10m above ground
u_component_of_wind_10m_above_ground
m/s
V component of wind 10m above ground
v_component_of_wind_10m_above_ground
m/s
Appendix
Supplementary Table 5: Processed GFS variables.
Site
Model
RMSE [W m -2 ]
FS [%]
rRMSE
Ramp recall [%]
MAE [W m -2 ]
MBE [W m -2 ]
Akwatia
Persistence
126.5
0.0
0.291
8.7
69.2
0.7
MLP
121.2
4.2
0.278
5.0
77.7
-14.3
SUNSET
120.1
5.1
0.276
0.6
82.2
-11.2
OmniVision
130.8
-3.4
0.3
5.0
79.0
-24.0
ViT
127.6
-0.8
0.293
5.0
80.6
-15.1
Golden
Persistence
107.5
0.0
0.209
7.1
45.1
2.7
Appendix
Supplementary Table 6: Benchmark performance of the forecasting models at a 10-min lead time. The best-performing value(s) for each site and evaluation metric are highlighted in bold . Ramp recall is evaluated using a 30% relative error threshold.
Site
Model
RMSE [W m -2 ]
FS [%]
rRMSE
Ramp recall [%]
MAE [W m -2 ]
MBE [W m -2 ]
Akwatia
Persistence
144.5
0.0
0.327
5.1
86.8
2.3
MLP
154.2
-6.8
0.349
4.5
113.6
-28.2
SUNSET
138.4
4.2
0.313
3.2
102.7
-12.9
OmniVision
149.0
-3.1
0.337
2.6
90.8
-15.9
ViT
136.9
5.2
0.31
2.6
88.4
-8.6
Golden
Persistence
146.2
0.0
0.281
6.8
65.6
8.6
Appendix
Supplementary Table 7: Benchmark performance of the forecasting models at a 30-min lead time. The best-performing value(s) for each site and evaluation metric are highlighted in bold . Ramp recall is evaluated using a 30% relative error threshold.
Site
Model
RMSE [W m -2 ]
FS [%]
rRMSE
Ramp recall [%]
MAE [W m -2 ]
MBE [W m -2 ]
Akwatia
Persistence
173.5
0.0
0.385
10.1
110.5
3.7
MLP
210.1
-21.1
0.466
3.4
165.9
-42.7
SUNSET
166.9
3.8
0.37
0.7
132.5
-10.7
OmniVision
168.1
3.1
0.373
9.4
121.4
-7.8
ViT
156.7
9.7
0.347
1.3
112.1
-22.2
Golden
Persistence
172.4
0.0
0.326
1.7
83.6
17.2
Appendix
Supplementary Table 8: Benchmark performance of the forecasting models at a 60-min lead time. The best-performing value(s) for each site and evaluation metric are highlighted in bold . Ramp recall is evaluated using a 30% relative error threshold.
Site
Model
RMSE [W m -2 ]
FS [%]
rRMSE
Ramp recall [%]
MAE [W m -2 ]
MBE [W m -2 ]
Akwatia
Persistence
175.0
0.0
0.38
5.5
118.8
3.3
MLP
245.8
-40.4
0.534
0.7
196.9
-53.8
SUNSET
185.2
-5.8
0.403
2.1
150.9
-15.9
OmniVision
208.6
-19.2
0.454
3.4
161.0
-40.1
ViT
182.7
-4.4
0.397
2.1
136.4
-3.4
Golden
Persistence
194.2
0.0
0.358
1.7
99.7
27.9
Appendix
Supplementary Table 9: Benchmark performance of the forecasting models at a 90-min lead time. The best-performing value(s) for each site and evaluation metric are highlighted in bold . Ramp recall is evaluated using a 30% relative error threshold.
Site
Model
RMSE [W m -2 ]
FS [%]
rRMSE
Ramp recall [%]
MAE [W m -2 ]
MBE [W m -2 ]
Akwatia
Persistence
191.8
0.0
0.411
6.2
137.1
3.3
MLP
288.3
-50.3
0.617
1.4
233.8
-67.4
SUNSET
201.5
-5.0
0.431
0.7
160.5
-6.4
OmniVision
205.7
-7.2
0.44
4.2
154.7
-1.1
ViT
199.3
-3.9
0.427
3.5
147.9
15.3
Golden
Persistence
205.4
0.0
0.372
0.0
110.8
38.9
Appendix
Supplementary Table 10: Benchmark performance of the forecasting models at a 120-min lead time. The best-performing value(s) for each site and evaluation metric are highlighted in bold . Ramp recall is evaluated using a 30% relative error threshold.
Solar power forecasting predicts photovoltaic generation and irradiance from weather, imagery, plant data, and recent output to support reliable grid and energy decisions.
Due to the rise in the use of renewable energies as an alternative to traditional ones, and especially solar energy, there is increasing interest in studying how to address photovoltaic forecasting in the face of the challenge of variability in photovoltaic energy production, using different methodologies. This work develops a hybrid approach for short and long-term forecasting based on two studies with the same purpose. A multimodal approach that combines images of the sky and photovoltaic energy history with meteorological data is proposed. The main goal is to improve the accuracy of ramp event prediction, increase the robustness of forecasts in cloudy conditions, and extend capabilities beyond nowcasting, to support more efficient operation of the power grid and better management of solar variability. Deep neural models are used for both nowcasting and forecasting solutions, incorporating individual and multiple meteorological variables, as well as an analytical solar position. The results demonstrate that the inclusion of meteorological data, particularly the surface long-wave, radiation downwards, and the combination of wind and solar position, significantly improves current predictions in both nowcasting and forecasting tasks, especially on cloudy days. This study highlights the importance of integrating diverse data sources to improve the reliability and interpretability of solar energy prediction models.
Ines Montoya-Espinagosa, Antonio Agudo
Universitat Pompeu Fabra, Barcelona, Spain. · Institut de Robòtica i Informàtica Industrial, CSIC-UPC, Barcelona, Spain.
Accurate solar irradiance forecasting is essential for the reliable integration of photovoltaic power into modern electricity grids. All-sky imagers (ASI) provide high-resolution observations of clouds, making them well suited for intra-hour forecasting. Recent deep learning approaches have substantially improved forecast accuracy but are often limited by deterministic predictions and a reduced capability to anticipate ramp events. This work proposes FarSky, a generative forecasting framework that leverages latent-space coupling to learn task-aware representations of sky images. A multi-task autoencoder first learns a shared latent representation for image reconstruction and irradiance estimation. A latent diffusion model then generates future latent states conditioned on recent observations, from which irradiance forecasts are directly decoded. Probabilistic forecasts are inherently obtained through stochastic sampling. The framework is developed using a multi-year ASI dataset acquired at the Plataforma Solar de Almería, Spain, and evaluated on two independent test datasets against persistence, state-of-the-art end-to-end, and generative forecasting approaches. FarSky achieves the best overall deterministic and probabilistic forecasting performance, improving forecast skill by up to 11 percentage points. Furthermore, it substantially improves ramp event detection over existing methods, achieving F1-scores above 60%. These results demonstrate the potential of combining generative models with task-aware latent-space coupling for solar forecasting.
Yann Fabel, Bijan Nouri, Milon Miah +4
1DLR, Institute of Solar Research · 2RWTH Aachen University · 3CIEMAT, Energy Department, Renewable Energy Division
Solar panels are increasingly deployed in cities on rooftops, walls, and urban infrastructure. Although the panel costs have fallen in recent years, the soft costs of installing them have not. These soft costs include assessing the illumination (irradiance) of a panel, which is typically performed using a 3D model that fails to capture small nearby structures that impact the irradiance. Our approach uses a single image taken at the panel's location to forecast its irradiance at any time in the future. We use visual cues in the image to find the camera's orientation and the portion of the sky visible to the panel in order to forecast the irradiance due to the sun and the sky. In addition, we show that the irradiance due to reflections from nearby buildings varies smoothly over time and can be forecasted from the image. This approach enables assessing the solar energy potential of any surface and forecasting the temporal variation of a panel's irradiance. We validate our approach using real irradiance measurements in urban canyons. We show that our approach often yields more accurate irradiance forecasts compared to conventional irradiance-based transposition methods and 3D model-based simulations. We also show that a single spherical image can be used to find the best fixed orientation of a panel. Finally, we present Solaris, a device to capture the image seen by a panel in a variety of urban settings.
Jeremy Klotz, Shree K. Nayar
Computer Science Department, Columbia University, 500 West 120th St., New York, 10027, NY, USA