The revolution of smart city has led to rapid development and proliferation of Internet of Things (IoT) technologies, with the focus on transmitting raw sensory data into valuable knowledge. Meanwhile, the ubiquitous deployments of IoT are raising the importance of processing data in real-time at the edge of networks rather than in remote cloud data centers. Based on above, edge computing has been proposed to exploit the capabilities of edge devices in providing in-proximity computing services for various IoT applications. In this paper, we present UrbanEdge, a conceptual edge computing architecture empowered by deep learning for urban IoT time series prediction. We design a hierarchical architecture to process correlated IoT time series and illustrate the work-flow of UrbanEdge in data collection, data transmission and data processing. As a core component of UrbanEdge, a deep learning model is developed with attention-based recurrent neural networks. Composed with multiple processing layers, the deep learning model can extract feature representations from raw IoT data for monitoring and prediction. We evaluate the designed deep learning model of UrbanEdge on real-world datasets, evaluation results show that the UrbanEdge outperforms other baseline methods in time series prediction.