The comparison between two different algorithms of spatio-temporal forecasting for traffic flow prediction

Abstract

Nowadays, there is an extensive body of literature that demonstrates the methods of forecasting traffic flows, which includes artificial neural networks, Kalman filtering, support vector regression, (seasonal) ARIMA models. However, seldom articles use two or more than two methods to predict the traffic flows and compare their difference within the forecasting process, which might be gradually recognized as a potentially important research area in the future. Two of the most commonly adopted methods, Space-Time Autoregressive Integrated Moving Average (STARIMA) and the Elman Recurrent Neural Network (ERNN), an Artificial Neural Network, have been firstly harnessed to establish the space-time predicting models. Secondly, according to the successfully trained models, the dissertation conducts the multi-dimensional comparison based on four aspects: interpretability; ease of implementation; running time and instability. Finally, some possible improvements are put forward according to their forecasting performance which also indirectly reflects their unique features and application environments.