Spatio-Temporal Recurrent Convolutional Networks for Citywide Short-term Crowd Flows Prediction

Abstract

With the rapid development of urban traffic, forecasting the flows of crowd plays an increasingly important role in traffic management and public safety. However, it is very challenging as it is affected by many complex factors, including spatio-temporal dependencies of regions and other external factors such as weather and holiday. In this paper, we proposed a deep-learning-based approach, named STRCNs, to forecast both inflow and outflow of crowds in every region of a city. STRCNs combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network structures to capture spatio-temporal dependencies, simultaneously. More particularly, our model can be decomposed into four components: Closeness captures the changes of instantaneous flows; Daily influence detects the changes of daily influence flows regularly; Weekly influence reacts weekly patterns of influence flows and External influence gets the influence of external factors. For the first three properties (Closeness, Daily influence and Weekly influence), we give a branch of recurrent convolutional network units to learn both spatial and temporal dependencies in crowd flows. External factors are fed into a two-layers fully connected neural network. STRCNs assigns different weights to different branches, and then merges the outputs of the four parts together. Experimental results on two data sets (MobileBJ and TaxiBJ) demonstrate that STRCNs outperforms classical time series and other deep-learning-based prediction methods.