Short-term prediction of traffic flow based on gated recurrent unit neural networks

Traffic management plays an important role in transportation planning, and an effective means of traffic management is traffic flow prediction. Traffic flow prediction provides traffic decision-makers the opportunity to reach accurate information to make judicious decisions that can reduce traffic congestion. On the other hand, deep neural networks promise to predict traffic flow accurately using big data driven from road detectors. This paper applied a model based on Gated Recurrent Units (GRU) to predict traffic volume by considering the time dependency of the time-series predicted variable. The hourly volume data are obtained from the Poledokhtar-Andimeshk rural road detector as well as time and holiday-related explanatory variables. The dataset includes ۱۶۸۰ hourly data; ۸۰% and ۲۰% considered for train and test set respectively. The model compiled using two loss functions, mean square error (MSE) and mean absolute error (MAE), and Adam optimizer. The results based on the MSE loss function show that loss can be reached near ۰.۱۶ after ۱۰۰ epochs and the results based on MAE show that loss can be reached near ۰.۳ after ۱۰۰ epochs that these results confirm GRU neural network models are powerful tools in traffic flow prediction.

کلیدواژه ها:

Short term traffic flow prediction ، deep learning ، GRU ، time series prediction