With the rapid proliferation of smart grid technologies, a large amount of fine-grained power data records have been collected and stored by different parties (e.g., the power supply bureau). Pooling together the records held by the parties makes mining the data value, and therefore promises enhanced energy management and efficiency. Despite the benefits of sharing these data, it also raises concerns about data privacy and security. To this end, we present a novel approach for privacy-preserving cross-party power data sharing approach in light of the Generative Adversarial Network (named PowerGAN), which enables the involved parties to construct a shared dataset without compromising the privacy of these parties. In PowerGAN, a centralized curator is assigned a generator, while each party possesses a discriminator. The key idea of PowerGAN is to let data holders jointly train the generator held by the centralized server. In addition, to prevent the curator from inferring sensitive data about the parties, we designed a privacy-preserving RMSProp (Root Mean Square Propagation) optimizer. Furthermore, we design a dynamic noise perturbation method, which dynamically tunes the noise to further promote the utility of the final shared data. Through comprehensive privacy analysis, we show that our PowerGAN approach provides strict privacy protection. Evaluations of real-world datasets show the effectiveness of PowerGAN in addressing the privacy concerns associated with multi-party power data sharing.