For comprehensive energy application by thermal power plants, the heating steam caused by the heating steam dispatching control and pipeline delay does not meet, in timely fashion, the needs of heating users. From the length of time delay, it is proposed to replace the real-time heating demand with the predicted value of heating user demand as the constraint condition of heating dispatch. Based on Informer, a multi-step time series prediction model of heat load is established. The prob-sparse self-attention mechanism adopted in the Informer model can effectively obtain the information in the heat load time series data, and establish the nonlinear relationship between heat load and related meteorological factors, thus improving the accuracy of heat load prediction. The verification of actual field data shows that the multi-step time series prediction model of heat load based on Informer established in this paper can have a strong fit with the heat load in a future period of time, and meet the demand of actual scheduling control delay for heat load.