Elbaz Mostafa, Said Wael, Mahmoud Gamal M, Marie Hanaa Salem
Department of Computer Science, Faculty of Computers and Informatics, Kafrelsheikh University, Kafrelsheikh, Egypt.
Computer Science Department, Faculty of Computers and Informatics, Zagazig University, Zagazig, 44511, Egypt.
Sci Rep. 2025 May 13;15(1):16635. doi: 10.1038/s41598-025-00600-7.
Integrating energy and solar imagery is essential for electrical engineers in renewable energy prediction, consumption analysis, regression modeling, and fault detection applications. A significant challenge in these areas is the limited availability of high-quality datasets, which can hinder the accuracy of the predictive models. To address this issue, this paper proposes leveraging Generative Adversarial Networks (GANs) to generate synthetic samples for training. Despite their potential, traditional GAN face challenges such as mode collapse, vanishing gradients, and pixel integrity issues. This paper introduces a novel architecture, Penca-GAN, which enhances GANs through three key modifications: (1) dual loss functions to ensure pixel integrity and promote diversity in augmented images, effectively mitigating mode collapse and improving the quality of synthetic data; (2) the integration of an identity block to stabilize training, preserving essential input features and facilitating smoother gradient flow; and (3) a pancreas-inspired metaheuristic loss function that dynamically adapts to variations in training data to maintain pixel coherence and diversity. Extensive experiments on three renewable energy datasets-SKY images, Solar images, and Wind Turbine images-demonstrate the effectiveness of the Penca-GAN architecture. Our comparative analysis revealed that Penca-GAN consistently achieved the lowest Fréchet Inception Distance (FID) scores (164.45 for SKY, 113.54 for Solar, and 109.34 for Wind Turbine), indicating superior image quality compared to other architectures. Additionally, it attains the highest Inception Score (IS) across all datasets, scoring 71.43 for SKY, 87.65 for Solar, and 90.32 for Wind Turbine. Furthermore, the application of Penca-GAN significantly enhanced the fault detection capabilities, achieving accuracy improvements from 85.92 to 90.04% for solar panels and from 86.06 to 90.43% for wind turbines. These results underscore Penca-GAN's robust performance in generating high-fidelity synthetic images, significantly advancing renewable energy applications, and improving model performance in critical tasks such as fault detection and energy prediction.
将能量与太阳能图像相结合，对于电气工程师在可再生能源预测、能耗分析、回归建模及故障检测应用中至关重要。这些领域的一个重大挑战是高质量数据集的可用性有限，这可能会阻碍预测模型的准确性。为解决这一问题，本文提出利用生成对抗网络（GAN）来生成用于训练的合成样本。尽管传统GAN有其潜力，但面临诸如模式崩溃、梯度消失和像素完整性问题等挑战。本文介绍了一种新颖的架构Penca - GAN，它通过三项关键修改来增强GAN：（1）双损失函数，以确保像素完整性并促进增强图像的多样性，有效减轻模式崩溃并提高合成数据的质量；（2）集成恒等块以稳定训练，保留基本输入特征并促进更平滑的梯度流；（3）一种受胰腺启发的元启发式损失函数，它能动态适应训练数据的变化以保持像素连贯性和多样性。在三个可再生能源数据集——天空图像、太阳能图像和风力涡轮机图像上进行的大量实验证明了Penca - GAN架构的有效性。我们的比较分析表明，Penca - GAN始终获得最低的弗雷歇因ception距离（FID）分数（天空图像为164.45，太阳能图像为113.54，风力涡轮机图像为109.34），表明与其他架构相比图像质量更优。此外，它在所有数据集中获得最高的Inception分数（IS），天空图像得分为71.43，太阳能图像为87.65，风力涡轮机图像为90.32。此外，Penca - GAN的应用显著增强了故障检测能力，太阳能电池板的准确率从85.92%提高到90.04%，风力涡轮机的准确率从86.06%提高到90.43%。这些结果强调了Penca - GAN在生成高保真合成图像方面的强大性能，显著推进了可再生能源应用，并在故障检测和能源预测等关键任务中提高了模型性能。
