Online-Offline Interactive Urban Crowd Flow Prediction Toward IoT-Based Smart City

Abstract

Urban crowd flow prediction is very challenging for public management and planning in smart city applications. IoT based technologies make urban-scale flow detection and prediction possible. Existing work mostly focuses on spatial and temporal dependence based flow prediction by learning patterns from historical crowd flow data with prior knowledge such as weather, events and location attributes, etc. However, these approaches are not well suited for predictions of instantaneous flow change usually due to social emergency incidents and accidents, which are not with obvious patterns but vital for urban safety. In this article we propose an Online to Offline Interaction based Dilated Casual Convolutional Neural Network framework (O2O-DCNN) to make predictions on urban crowd flow. Both online attention behavior and offline crowd shift factors are considered in our framework, in case to capture the dependence between them and make more accurate predictions especially for instantaneous flow variations. The online and offline features are processed by dilated casual convolutions and then put into CBOW model based full connected network to make interactions. Our framework combines the causality of tempo-spatial related flow time series and semantic-based online attention behavior time series without too deep layers of neural network. The performance evaluations are based on realistic User Detail Record (UDR) dataset of a southern city in China provided by China Unicom. O2O-DCNN is compared with the other related baselines in terms of MASE and MAE. The results show that our framework is with much better accuracy, especially for instantaneous flow variation scenarios.