Customized Parking Data Generation based on Multi-conditional GAN

Parking data is vulnerably affected by spatiotemporal characteristics and surrounding societal events, causing the latent features of the parking data are hard to learned by GANs if solely given the time-series parking data. Hence it is impossible to generate the desired data with high quality. In this paper, we propose a multi-conditional GAN, named MCGAN to refine the generating process and optimize the generating quality via introducing external customized extendable conditions related to the parking data samples. These conditions, in forms of condition tensors in MCGAN, can help the network learn the features introduced by each defined condition and will reproduce, even combine them in the later generating process, achieving the better result. The experiments show the working process of MCGAN is not different very much from GANs, but the generating quality get improved greatly if given the output expectation more specifically.