Abstract: The physical model of distribution networks exhibits significant non-convex nonlinearities, making it challenging to achieve coordinated optimization of active and reactive power. To address this issue, this paper proposes a coordination optimization strategy for active and reactive power in distribution networks that does not rely on physical models. Firstly, an initial data-driven model for distribution networks is established based on a convolutional neural network (CNN). Secondly, the progressive adversarial cycle network (PACN) technology, which incorporates a generator and a binary classification module, is introduced. On the one hand, pseudo-measurement data is obtained through the generator to augment the dataset; on the other hand, the binary classification module screens out high-quality pseudo-measurement datasets. Thirdly, the dataset generated by the PACN is applied to the initial data-driven model of the distribution network to enhance its accuracy. Furthermore, a correction term is added to the model to penalize unknown states during the operation of the distribution network, thereby improving its network adaptability. Finally, the proposed method is validated using the IEEE 33-bus system, and the results demonstrate that the method effectively enhances the system's security and economy.