Short-term traffic flow prediction with LSTM recurrent neural network

Accurate and timely short-term traffic flow prediction plays an important role in intelligent transportation management and control. Traffic flow prediction has a long history and is still a difficult problem due to intrinsically highly nonlinear and stochastic characteristics of complex transportation systems. In this paper, we employ the long short-term memory (LSTM) recurrent neural network to analyze the effects of various input settings on the LSTM prediction performances. Flow, speed, and occupancy at the same detector station are used as inputs to predict traffic flow. The results show that the inclusion of occupancy/speed information may help to enhance the performance of the model overall. Further, we include as inputs traffic variables from the upstream and/or downstream detector stations for traffic flow prediction. The evaluation of such spatial-temporal input interactions show that the inclusion of both downstream and upstream traffic information is useful in improving prediction accuracy.