Predicting Traffic Flow Based on Encoder-Decoder Framework

Predicting traffic flow is of great importance to traffic management and public safety, and it has high requirements on accuracy and efficiency. However, the problem is very challenging because of high-dimensional features, spatial levels, and sequence dependencies. On the one hand, we propose an effective end-to-end model, called FedNet, to predict traffic flow of each region in a city. First, for the temporal properties, we obtain low-dimensional features by downsampling high-dimensional input features. Then we perform temporal fusion to get temporal aggregations of different spatial levels. Next, we generate traffic flow by upsampling the fused features which are obtained by combining the corresponding temporal aggregation and the output of the previous upsample block. Finally, the traffic flow is adjusted by external factors like weather and date. On the other hand, we transfer the original task into a sequence task and then use teacher forcing to train our model, which make it learn the sequence dependencies. We conduct extensive experiments on two types of traffic flow (new-flow/end-flow and inflow/outflow) in New York City and Beijing to demonstrate that the FedNet outperforms five well-known methods.