AI for Science
xGBoost outperformed other machine learning models at forecasting rooftop solar power in a Saudi commercial building
In a 2025 paper in Energy Reports, researchers forecast rooftop photovoltaic solar generation at a commercial building in Saudi Arabia. They tested several machine learning models on weather and power data gathered at the site, including module temperature, ambient temperature, irradiance, and real power. The xGBoost model, short for Extreme Gradient Boost Regression, produced the most accurate forecasts among the models compared. The paper reports a mean squared error of 0.0036, a root mean squared error of 0.0191, a mean absolute percentage error of 0.6898, a mean absolute error of 0.0130, and an R-squared of 0.975. Lower error values indicate closer predictions, and a higher R-squared indicates a stronger fit.
What did the researchers test?
The authors write that accurate solar power prediction is essential for designing and managing solar power plants. The authors note that relying too heavily on solar power can strain grid management and power quality. The study used a time-series dataset of solar power from a photovoltaic system at a commercial building in Saudi Arabia.
What did the comparison show?
The authors evaluated the forecasting power of different machine learning approaches with several statistical indices. Among them, xGBoost had the best score on the reported measures. The authors write that xGBoost can produce accurate forecasts with low error and that it outperforms several other models in forecasting accuracy.
FAQ
What data did the study use?
The study used measurements from a photovoltaic system at a commercial building in Saudi Arabia. The data were module temperature, ambient temperature, irradiance, and real power.
Which model did best?
The paper reports that xGBoost, also called Extreme Gradient Boost Regression, did best. It reports an R-squared of 0.975 and the error values listed above.
Why does prediction matter?
The authors write that better solar forecasts help the distribution grid run more smoothly and support the design and management of solar plants.
Limits and what's next
The study examined a photovoltaic system at a commercial building in Saudi Arabia. The researchers based the forecasts on data collected at that site, including module temperature, ambient temperature, irradiance, and real power. The abstract reports no independent test at another site. The authors also note that machine learning work on solar power in Saudi Arabia is limited compared with advances elsewhere. This is one study, and the abstract reports no independent replication at another location.
Source: Upma Singh et al., “Forecasting rooftop photovoltaic solar power using machine learning techniques,” Energy Reports, 2025.