The real-time and precise short-term traffic flow prediction is the key factor for the realizing of traffic control and traffic guidance in the intelligent traffic system. As the complexity of short-term traffic flow data, the saturated correlation dimension method and Cao’s method are adopted to calculate the embedding dimension and delay time of traffic flow time series, and the Wolf method is applied to calculate the largest Lyapunov exponent of the reconstructed traffic flow time series. The result shows that the traffic flow series is a chaotic sequence with better predictability. Then the prediction methods based on ESN and Elman neural networks are applied to predict traffic flow time series respectively, it indicates that the former has quicker prediction speed on the condition of the same prediction accuracy.