Traditional data restoration methods are limited in accuracy for distribution transformer area load data because they neglect external spatiotemporal correlations. This paper proposes a load data restoration method that combines reference values and an improved generative adversarial network (GAN). First, a reference value construction method is developed, which integrates external information with the load’s own spatiotemporal correlations to generate an initial restoration point that approximates the true data distribution. Then, within the GAN structure, dynamic positional encoding is designed to replace random noise, enabling the model to capture contextual information of missing data. Hybrid residual dilated convolutions and a channel attention mechanisms are embedded to enhance the model’s capability in extracting load fluctuation features. In addition, a multi-scale structural similarity loss is introduced into the original loss function to optimize the training performance. Finally, validation based on real measured data shows that the proposed method improves restoration accuracy by approximately 40% compared to traditional data restoration methods.