This paper proposes an automated configuration optimization algorithm based on Wasserstein Generative Adversarial Networks (WGAN) called COGAN, aimed at efficiently tuning IoT device configurations in power grid systems under resource constraints in various environments. Unlike traditional methods that rely on manual adjustments or large sample prediction models, COGAN uses adversarial training to generate optimized configurations that meet specific resource limits (e.g., CPU, memory). The generator learns the distribution patterns of high-performance configurations, while the discriminator ensures the quality and stability of the generated configurations. Experimental results show that the configurations generated by the COGAN algorithm outperform the default configurations, providing an effective solution for adaptively optimizing IoT device configurations in real-world deployment environments, enhancing the plug-and-play capability of devices.