September 21, 2022
Evaluating Performance of Different Generative Adversarial Networks for Large-Scale Building Power Demand Prediction
AbstractUtilizing the advantages of both physical-based and data-driven methods, hybrid models are often used to predict building power demand at a large scale. Generative Adversarial Networks (GAN), one data-driven method, recently attracts a lot of attention for the hybrid models. This paper evaluates the performance of three types of GANs (original GAN, cGAN, and ACGAN) for predicting building power demand at a large scale. First, physical-based building models are developed and simulated by using EnergyPlus. Then, power demands for a partial set of buildings simulated by EnergyPlus are used to train the GAN models. In the performance evaluation of different GANs, power demands for all buildings simulated by EnergyPlus are used as a reference. There are three main findings: (1) Original GAN and cGAN have better performance than ACGAN; (2) Original GAN has good performance even if the number of training samples is limited; and (3) cGAN has better performance when the number of building types increases. Due to the advantages of the hybrid models and GAN, a hybrid model using GAN is a promising way for a study of building power demand prediction at a large scale.
Published: September 21, 2022