Daily power demand prediction for buildings at a large scale using a hybrid of physics-based model and generative adversarial network

Power demand prediction for buildings at a large scale is required power grid operation. The bottom-up method using physics-based models popular, but has some limitations such as heavy workload on model creation and long computing time. Top-down methods based data driven are fast, less accurate. Considering the similarity of patterns single superiority generative adversarial network (GAN), this paper proposes new (E-GAN), which combines (EnergyPlus) data-driven to predict daily scale. E-GAN selects small number typical utilizes EnergyPlus their demands. Utilizing those buildings, GAN then adopted forecast demands buildings. To verify proposed method, used 24-hour set residential results show that (1) 4.3% in each building category ensure accuracy; (2) compared with model, can accurately only 5% error (measured by mean absolute percentage error, MAPE) while approximately 9% time; (3) (e.g., support vector regression, extreme learning machine, polynomial regression model), demonstrates least 60% reduction measured MAPE.