Abstract：

This paper conducts a study on the data generative model of heat supply stations. A data generation model is established based on the generative adversarial network (GAN) and the denoising diffusion probabilistic model (DDPM). By learning the joint distribution of meteorological, room temperature, and operational data of the heat supply station, the original training data is augmented to provide sufficient data support for the training of the forecast model, thereby improving the accuracy of the forecast model. It has been verified and tested at a heat supply station in Beijing. The results show that this method can reduce the forecast errors of the opening degree of the primary side electric control valve of the heat supply station and the supply water temperature of the secondary network by about 7% and 11%, respectively. Meanwhile, the deviation between the expected room temperature obtained by adjusting the heat supply using the load forecast value with improved application accuracy and the target room temperature can be further reduced by 5.44%. The generative model based on GAN can expand the forecast range of the forecast model, and the generation model based on DDPM can improve the accuracy of the forecast model within the original forecast range. This study can provide support for further improving the load forecast ability and on-demand precise regulation level of heat supply stations in district heating systems.

Keywords:district heating; heat supply station; load forecast; data augmentation; generative adversarial network; denoising diffusion probabilistic model; generative model