Abstract: To deal with the deficiencies of the existing prediction models based on dynamic recurrent neural network, considering the characteristics of the traffic flow itself such as complexity, nonlinearity and uncertainty, a short-term traffic flow prediction method using Elman neural network with variable gain was proposed. The real-time updates of the network were achieved by introducing a variable gain factor based on the real-time error analysis. The method was verified by the field measured data from Renmin Street, Changchun City. The results show that comparing with the existing Elman network based models the proposed method is superior on the aspects of convergence time and prediction accuracy.

Key words: engineering of communications and transportation, traffic flow prediction, dynamic recurrent neural network, variable gain factor