Energies Media
Along Egypt’s Red Sea coast, the semi-arid strip of Zafarana already hosts some of the country’s largest wind farms — and researchers believe its solar potential may be even greater. But harnessing that potential depends on a deceptively difficult problem: knowing, hours or weeks in advance, exactly how much sunlight will reach the panels.
A new study put seven competing algorithms to the test on four years of local meteorological data. The winner wasn’t the deep learning architecture most researchers would have picked.
Predicting how much sunlight will reach a solar panel sounds straightforward. It isn’t. Temperature, humidity, wind speed, cloud cover, and time of day all shift constantly — and they interact in ways that resist simple modeling.
Getting those predictions wrong carries real costs. Underestimating output can force grid operators to keep fossil-fuel backup plants running unnecessarily, while overestimating it can leave the grid short when demand peaks. Either way, poor forecasts erode the economic case for solar investment.
Zafarana offers an unusually clean test environment. Its semi-arid climate and minimal cloud cover cut some of the noise that complicates forecasting elsewhere, but simpler conditions don’t mean a trivial problem. Irradiance still swings dramatically between dawn, midday, and dusk, and across seasons. To capture that complexity, researchers drew on four years of hourly data from NASA’s Power Project, covering 2020 through 2023 — more than 35,000 records in total.
Before training a single model, the team ran three separate feature-selection analyses on nine meteorological and temporal variables. The methods — One-Way ANOVA, Boruta, and Random Forest importance scoring — each approach the problem differently, and they didn’t always agree.
One-Way ANOVA flagged temperature, humidity, and wind speed as the most critical predictors, ranking the hour of day as relatively unimportant. That conclusion would later prove misleading. Random Forest importance scoring told a different story, placing the hour of day at the top — a ranking that reflects how sharply irradiance rises and falls across a 24-hour cycle.
Boruta dismissed the “year” variable as irrelevant, and that finding held up: removing it from classical models barely moved the needle on performance. Ablation tests confirmed the hour feature mattered most. Dropping it caused the sharpest accuracy drop across nearly all models, directly contradicting ANOVA’s earlier ranking.
With features identified, the researchers benchmarked seven algorithms: Linear Regression, Decision Tree, Gradient Boosting, Random Forest, CNN, LSTM, and a hybrid CNN-LSTM. Each was tested on both a one-day-ahead and a one-month-ahead forecasting horizon.
Random Forest came out on top by a clear margin. It achieved an R² score of 0.9948 for one-day-ahead forecasts and 0.978 for one-month-ahead forecasts — meaning it explained more than 97% of the variance in solar irradiance across both time horizons. Gradient Boosting followed closely, posting R² scores of 0.9724 for the short-term task and 0.954 for the long-term one.
The deep learning models — CNN, LSTM, and CNN-LSTM — scored notably lower on R². Researchers attribute this partly to dataset size. Neural networks are data-hungry, and 35,000 records may not give deep learning architectures enough signal to learn complex temporal dependencies reliably.
When the researchers ran paired t-tests comparing prediction errors directly, the rankings shifted. LSTM and CNN-LSTM produced significantly smaller absolute prediction errors than Random Forest — p-values below 0.0001 in both cases. At individual time steps, the deep learning models were more precise, even though they explained less overall variance.
That distinction has practical implications. R² measures how well a model tracks broad irradiance trends over time; mean absolute error measures how far off each individual prediction is. A grid operator planning daily energy budgets might prioritize R² and lean toward Random Forest. One managing real-time dispatch decisions might prefer LSTM’s tighter pointwise accuracy — a genuinely different use case, not just a preference.
The authors note that deep learning performance could improve substantially with larger datasets or more thorough hyperparameter tuning. The current finding reflects a specific dataset size and a deliberately simple architecture, not a ceiling on what deep learning can achieve at Zafarana.
The conditions that make Zafarana a useful test site — high irradiance, semi-arid climate, minimal cloud cover — are shared by large parts of North Africa, the Middle East, and other sun-rich developing regions. The study’s findings carry potential relevance well beyond Egypt’s Red Sea coast.
One practical advantage of the framework is its data source. NASA’s Power Project data is freely available online, meaning the same approach could be replicated at low cost in regions where ground-based meteorological monitoring is sparse or unreliable. Random Forest also runs efficiently on standard hardware, no GPU clusters required — making it accessible to utilities and planners who lack large-scale computing infrastructure.
More broadly, the study adds to a growing body of evidence that well-tuned classical models can match or outperform deep learning when training data is limited. As more solar projects come online in data-scarce environments, that’s a finding worth watching.
The complete study can be found here: Taha, A., Makeen, P. & Nazih, N. Short-term and long-term solar irradiance forecasting with advanced machine learning techniques in Zafarana, Egypt. Sci Rep 15, 39553 (2025). https://doi.org/10.1038/s41598-025-24853-4
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked at international companies such as Siemens and is multilingual.
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked at international companies such as Siemens and is multilingual.
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked at international companies such as Siemens and is multilingual.

source