spatial-temporal prediction in cyber-physical systems
Friday, June 04, 2021, 08:00am - 05:00pm
Speaker: Shuxin Zhong
Location : Join Zoom Meeting
Prof.Desheng Zhang, Prof.Yongfeng Zhang, Prof.Hao Wang, and Prof.Jie Gao.
Event Type: Qualifying Exam
Abstract: Spatial-temporal predictions are essential to Cyber-Physical Systems (CPS) such as the taxi network and bus network to satisfy users’ demand. In order to manage those CPS resources efficiently, it is vital to predicting the users’ demand ahead of time for future possible actions such as rescheduling.However, current methods, while working fine for short-term prediction, do not work well to predict the unexpected increasing user demand in advance (i.e., ahead of sufficient time, say 2 hours) and thus are inefficient for resultant applications such as bike re-balancing. To address this issue, we leverage the fact that the latest public transportation traffic demand as an example of large-scale CPS, e.g., subways, can reflect the future bike traffic demand. Specifically, we model downstream traffic, i.e., bikes, and upstream traffic, i.e., subways, as spatial-temporal network series, and then design a multi-modal 3D deep capsule network framework called BikeCap. The key novelty of BikeCap is in our first attempt to apply the capsule network in a temporal domain, which benefits the prediction of a long-term bike demand.BikeCap has three components: (1) a pyramid-based convolutional layer in historical capsules to capture the spatial-temporal correlations along the directions of traffic propagation; (2) a spatial-temporal routing technique to actively learn the uncertain multi-modal spatial-temporal correlations between historical capsules and future capsules; (3) a 3D deconvolution-based decoder to construct future bike demand considering the multi-dimensional correlations of traffic patterns among neighborhoods.Through experiments based on the data of 30,000 bikes and 7 subway lines collected in Shenzhen City, China, the results show that BikeCap outperforms several state-of-the-art methods. We also conduct an ablation study to show the impact of BikeCap’s different designed components and multi-modal concept.