The Cloudy Sky Advantage: How Sharper Satellites Improve Solar Forecasting — and Where They Still Fall Short

On a gray January morning in Estonia, a neural network stares at satellite images of clouds and makes a prediction: 187 watts per square meter of solar energy will hit the ground in the next ten minutes. It's wrong by 12 watts. That's actually impressive — and it represents a meaningful step forward in our ability to forecast solar power, a technology that now generates roughly 5% of the world's electricity and is growing faster than any other energy source.

The gap between "impressive" and "just okay" in solar forecasting depends heavily on what the sky is doing. When thick clouds dominate — when the atmosphere itself is doing most of the work in blocking sunlight — a new study shows that feeding a machine learning model sharper satellite images yields substantial gains. But when skies are clear and the sun dominates, even the sharpest imagery can't rescue neural networks from their blind spots. The study, published by researchers at the University of Tartu in Estonia, is one of the first to quantify exactly how much the next generation of European weather satellites matters — and where they hit a wall.

The Science

Meteosat Third Generation (MTG) represents the most significant upgrade to European weather satellite imaging in two decades. The first operational satellite launched in December 2022, and by late 2024, its primary imager — the Flexible Combined Imager, or FCI — began delivering data at resolutions far exceeding its predecessor. While the older Meteosat Second Generation (MSG) captured visible light at 3 kilometers per pixel at the center of its field of view (and a single high-resolution channel at 1 kilometer), MTG sees the world at 500 meters per pixel for certain visible channels, with additional infrared bands at 1 kilometer. That's four to six times more pixels covering the same ground.

For solar energy forecasting, this resolution difference could matter more than it might initially seem. Clouds don't have clean edges. A cloud bank that appears as a single block in a 3-kilometer image might actually have wispy edges, holes, or sub-structures that affect how much sunlight leaks through to the ground. If a model can see those edges, it might better predict exactly how much irradiance reaches a solar farm or a rooftop panel.

But nobody had rigorously tested whether this sharper view actually helps machine learning models estimate surface solar irradiance (SSI) — the amount of solar energy arriving at ground level. That's the gap Gordei Pribõtkin, Piia Post, and Velle Toll set out to fill.

Their approach combined two satellite systems over one country: Estonia, a small Baltic nation whose eight meteorological stations provided ground-truth measurements for five years of MSG data (2021–2025) and one year of MTG data (2025). The researchers built a convolutional neural network — a type of deep learning model particularly good at extracting patterns from images — that could ingest multiple image patches simultaneously. The architecture handled four different "modalities": MSG's 3-kilometer infrared bands, MSG's 1-kilometer high-resolution visible channel, MTG's 1-kilometer infrared bands, and MTG's 500-meter visible and near-infrared channels.

Rather than training two separate models and hoping for the best, the team took a transfer learning approach. They first trained a model on the five years of MSG data alone — a "SEVIRI-only" baseline that learned from an enormous dataset spanning all four seasons repeatedly. Then, they froze the MSG-processing parts of that network and fine-tuned only the new branches designed to handle MTG's sharper imagery. This "HYBRID" model inherited everything the network had learned from five years of data while learning to incorporate the additional detail from MTG.

The choice was practical and principled. Training any model on a single year of data tends to produce unstable results — the network might get lucky (or unlucky) depending on whether that particular year happened to be cloudier or clearer than average. By starting from the five-year model and only fine-tuning the new branches, the researchers could be more confident that any improvement came from MTG's resolution, not from statistical noise.

The test design was rigorous. Eight Estonian stations served as a geographical cross-validation set: in each fold, two stations were held out for validation (helping the model know when to stop training), and one station was held out for final testing. The model never saw any of these test locations during training. This spatial separation ensures the results aren't just the model memorizing specific sites.

Sky conditions — not just overall accuracy — drove the analysis. The researchers classified each ten-minute measurement into four categories: clear (daily variability close to clear-sky values and high overall irradiance), partly cloudy (high overall irradiance but more variability), cloudy (moderate daily irradiance with substantial cloud effects), and overcast (less than half of clear-sky irradiance, with clouds dominating). This distinction proved crucial to understanding where MTG's sharper vision helped and where it didn't.

What They Found

The headline result lands squarely on cloudy skies: the HYBRID model reduced root mean squared error by 8.2 watts per square meter under overcast conditions and 5.7 watts per square meter under cloudy conditions, compared to the SEVIRI-only model (Pribõtkin et al., 2026). Those aren't trivial improvements — under overcast skies, where the base error is high because clouds are unpredictable, cutting 8 watts from an RMSE of roughly 50–60 watts represents roughly a 13–16% reduction in forecasting error.

RMSE Reduction from Adding MTG/FCI Data by Sky Condition

Label	Value
Overcast	8.2 W/m²
Cloudy	5.7 W/m²
Partly Cloudy	0 W/m²
Clear	0 W/m²

Source: Pribõtkin et al. (2026). Error bars show standard error of the mean across cross-validation folds and random seeds.

The pattern becomes even clearer when comparing the HYBRID model to SARAH-3, the leading physics-based satellite product for Europe. The skill score — defined as one minus the ratio of model RMSE to baseline RMSE — quantifies how much better (or worse) a model performs than the reference. A skill score of 35% under overcast conditions means the HYBRID model made errors that were 35% smaller than SARAH-3's errors. Under cloudy conditions, the advantage was 21%; overall, 20%.

HYBRID Model Skill Score vs. SARAH-3 by Sky Condition

Label	Value
Overcast	35 %
Cloudy	21 %
Overall	20 %
Clear	-10 %

Source: Pribõtkin et al. (2026). Skill score = 1 - (RMSE_model / RMSE_baseline).

But the findings take a sharp turn in clear skies. Neither the SEVIRI-only model nor the HYBRID model could beat SARAH-3 when the sun shone unobstructed. Both machine learning approaches underperformed the physics-based baseline — and the difference wasn't trivial. The clear-sky regime, despite representing only 4.4% of samples in the test set, exposed a fundamental limitation in how these models learn.

The researchers offer two complementary explanations. First, under clear skies, the signal reaching the satellite comes from Earth's surface — the albedo, or reflectivity, of the ground. Surface albedo varies enormously: snow reflects differently than forests, cities reflect differently than oceans. A model trained in Estonia might struggle in Spain because the ground looks completely different. Under cloudy skies, the signal is dominated by clouds themselves, which look similar whether they're over Tallinn or Tenerife. Second, clear-sky irradiance depends strongly on atmospheric aerosols and water vapor — small particles and moisture that bend and absorb sunlight. These are notoriously hard to infer from visible and infrared satellite bands alone, and they have outsized effects precisely when skies are clearest.

The figure above shows the paired RMSE differences across sky conditions — each point represents one cross-validation fold, with error bars showing standard error of the mean across different random initialization seeds. The pattern is unmistakable: under overcast and cloudy conditions, nearly every fold favors the HYBRID model. Under partly cloudy and clear conditions, the differences scatter around zero — sometimes HYBRID wins, sometimes SEVIRI-only wins, and the average difference is indistinguishable from zero.

Why This Changes Things

Solar energy is no longer a niche technology awaiting its moment. In 2024, global solar photovoltaic capacity exceeded 1.5 terawatts — enough to power roughly 450 million average American homes. The International Energy Agency projects that solar will be the largest source of electricity generation worldwide by 2030, surpassing even natural gas. As this capacity grows, the importance of accurate forecasting grows with it.

The reason is grid stability. Electricity grids must match supply and demand moment to moment. A solar farm that produces 500 megawatts on a clear Tuesday in July and zero on a cloudy Wednesday in November isn't inherently a problem — as long as grid operators know it's coming.forecast error translates directly into financial risk and operational complexity. When a 200-megawatt solar plant underperforms its forecast by 20%, the grid must compensate by ramping up other generation, often natural gas plants that take time to start up and cost money to run hot. These "ramp events" become more expensive and more frequent as solar's share of the mix grows.

The results from Pribõtkin and colleagues suggest that the next generation of European satellites will meaningfully improve forecasts specifically during the most challenging conditions — when clouds are present and variability is high. This is precisely when today's forecasts struggle most and when the consequences of error are greatest. A 35% skill improvement over existing products under overcast conditions, if it holds in operational deployment, could translate to millions of euros in avoided balancing costs across Europe's solar fleet.

But the study's clear-sky finding is equally important, if more sobering. Machine learning models trained on satellite imagery alone — regardless of resolution — appear to hit a ceiling in conditions where physics-based models have inherent advantages. The Heliosat method underlying SARAH-3 encodes centuries of atmospheric physics: it explicitly models how sunlight interacts with clouds, aerosols, and gases. A neural network trained to mimic the outputs of that model can learn the patterns it has seen, but it struggles with the physical processes it hasn't observed enough of to learn.

This doesn't mean machine learning has no role in clear-sky forecasting — recent work has shown that incorporating additional data sources (reanalysis products, radiative transfer modeling) can close the gap. But it means that simply throwing sharper images at a neural network won't solve the problem. The researchers are explicit: "higher spatial resolution alone is insufficient to address clear-sky limitations in machine-learning-based SSI retrieval."

There's also a geographic caveat worth considering. The study was conducted over Estonia, which sits between 57.5° and 60.9° north latitude — considerably further from the equator than most of Europe's population and solar capacity. Northern Europe presents a unique combination of strong seasonal cycles, long twilight periods, and frequently overcast conditions. The finding that MTG's resolution matters most under cloudy conditions might be particularly pronounced in this region; the benefits could be smaller in sunny southern Europe where clear skies are more common and cloud-edge effects less dominant.

What's Next

The study opens several threads worth pulling. The first is geographic generalization: will these results hold in Spain, Greece, or North Africa, where solar farms are larger and clearer skies dominate? The authors acknowledge that their cross-validation tested locations within a small, climatically homogeneous country. Verbois et al. (2023) previously found that machine learning models struggle to generalize outside their training regions, and Pribõtkin et al. note that leakage from shared weather patterns likely inflates their test performance. A rigorous evaluation across Europe's climatic gradient is the obvious next step.

The second thread is temporal resolution. This study used 10-minute intervals — already a major improvement over the 30-minute resolution of SARAH-3 and most gridded products. But MTG's Flexible Combined Imager can deliver data every 2.5 minutes in its high-resolution mode. Faster updates could be critical for nowcasting — predicting irradiance minutes to an hour ahead, which is when the value of solar power forecasts is highest for grid operators.

Third, the study tested one specific architecture — a multi-branch ResNet-style convolutional network that processes each resolution separately before combining embeddings. Alternative architectures, particularly Vision Transformers that can model long-range spatial dependencies more flexibly, might exploit MTG's resolution differently. The field is young; the authors tested one reasonable approach, not the theoretically optimal one.

Finally, the clear-sky problem demands attention. If machine learning is to serve as a universal solar forecasting tool — working equally well in Germany's winter overcast as in California's summer clarity — researchers will need to find ways to inject physical constraints or auxiliary data into the training process. The authors suggest incorporating aerosol and water vapor data from external sources; others have proposed using physics-informed loss functions or hybrid architectures that combine neural networks with explicit radiative transfer modeling. None of these are simple, but the stakes are high enough that the effort seems justified.

The 2026 launch of Europe's next-generation satellite constellation was accompanied by promises of better weather forecasting, improved climate monitoring, and sharper environmental observation. This study offers the first concrete evidence of how those promises translate — or don't — for solar energy. In overcast skies above Estonia, the new eyes in orbit proved their worth. In clear skies above the same landscape, they reached a limit that no amount of pixel density can currently overcome. The path forward will require both sharper instruments and smarter models — and, likely, some creative combination of the two.