The current availability of real-time datasets from electronic devices in cities has offered an unprecedented opportunity to develop methods for controlling of traffic flow and implementation of intelligent transportation systems. Moreover, the mapping of city structure as networks has enabled to understand how the organization of streets influences the transport of people and goodies. In this work, we consider data from electronic radars in the Sorocaba city (Brazil) and constructed a graph, whose nodes are radars and two radars are connected according to geographical distances. We propose a model in which the number of cars passing at radar j at time (t +Δt) depends the flow on its neighbors. The model is evaluated through artificial neural networks and random forests. We predict the number of cars passing by a radar in an interval of length Δt = 30 minutes. Indeed, considering a two minutes resolution, our model can predict the number of cars with an average error close to two cars per minute. We verified that the closer the time interval accounted, the better the prediction. The method proposed here is general and can be used in any complex systems whose elements have time series associated.
