The AI model tightens the solar forecasts to support the stability of the satellite network
Accurate predictions of solar radiation are crucial for the stability of photovoltaic energy systems, but current models often fade as the prediction time increases. To tackle this challenge, researchers led by Nanjing University of Information Science and Technology have unveiled an AI-based solution called GAN-Solar, designed to generate sharper, more reliable forecasts for solar energy management.
The model uses the principle of Generative Adversarial Networks (GANs), pitting two neural networks against each other in a process that resembles a competition between a ‘master painter’ (the generator) and a ‘sharp art critic’ (the discriminator). The generator produces simulated future radiation maps based on historical data, while the discriminator learns to detect whether the images are real or generated.
“This continuous adversarial training continually hones the ‘painter’s’ skills, ultimately allowing him to produce accurate, high-resolution predictions that are almost indistinguishable from reality,” explains Chao Chen, lead author of the study published in the International Journal of Intelligent Networks.
Unlike conventional models that lose detail over time, GAN-Solar offers higher reliability in both the global distribution and local characteristics of solar radiation. “Traditional models ‘see’ less clearly over longer forecast times. GAN-Solar is like equipping the forecast system with high-precision glasses,” Chen said. The improved accuracy supports smoother operation of solar power grids and satellite networks that rely on stable energy input.
Experimental validation shows that GAN-Solar increased the Structural Similarity Index (SSIM) of predicted images from 0.84 to 0.87, outperforming other state-of-the-art models. The results demonstrate the company’s ability to deliver highly accurate, low-distortion forecasts that are essential for real-time solar applications and satellite communications networks.
Research report:GAN-based optimization of solar radiation prediction for satellite communication networks
