A Trade-off Between Accuracy and Complexity: Short-term Traffic Flow Prediction with Spatio-temporal Correlations

Considering spatio-temporal correlation between traffic in different roads has benefit for building an accurate spatio-temporal model for traffic prediction. However, it implies high computational complexity for model building in the context of a complicated network topology, e.g., urban network. Hence, this paper develops a method for capturing and quantifying the intricate spatio-temporal correlations. The contributions of this paper are as follows. First, we offer a physically intuitive approach to capture the spatio-temporal correlation between traffic in different roads, which is related to the road network topology, time-varying speed, and time-varying trip distribution. With this approach, only the parameters, namely time-varying lags, in our STARIMA (Space-Time Autoregressive Integrated Moving Average) based model should be adjusted in different time periods of the day. It guarantees the prediction accuracy and makes the predictor readily amendable to suit changing road and traffic conditions. Second, a metric named traffic transition probability calculated based on trip distribution, as well as a threshold ε are applied for selecting the most spatio-temporally correlated neighbors of a target road. Thus, the complexity of model building will be reduced dramatically. Trace-driven experiments are conducted from two aspects. First, our proposed prediction method has superior accuracy compared with ARIMA and the back propagation neural network model (BPNN) based method, but has much reduced computational complexity. Second, the results show that the prediction accuracy is not always proportional to the increase in the number of spatial neighbors considered for a target road. The trade-off between accuracy and complexity depends on the configuration of ε.