A Prediction Approach For Traffic Flow Based On Short-Long Term Combination Model

Getting the real-time information of traffic flows was one of the important steps to achieve the intelligent transportation. In the Intelligent Transportation Systems, the forecasting of dynamic traffic flow on the road was an essential part of the Traffic Monitoring System. The traffic flow conditions of specified sections were the necessary reference for congestion management, vehicle's dynamic route guidance, traffic management, traveler information system and incident detection system. A good traffic flow forecasting methods could provide true and reliable data for traffic control. Prediction of traffic flow was divided into short-term forecasting and long-term forecasting. The short-term forecasting had the real-time which made it adapt to the change of the real-time data easily. The long-term forecasting mainly reflected the regularity of traffic flow in a longer period of time, for the reason that not all state changes were permanent. This paper put forward the methods combinated the short-term and long-term prediction based on the existing prediction methods. It relied on the traffic projections, and used the data obtained from the Traffic Monitoring System to make validation of prediction methods.Firstly, this article used the double exponential smoothing for the short-term forecasting. In the double exponential smoothing forecasting, it used Levenberg-Marquardt algorithm to optimize the smoothing parameters so that the error of artificial selection parameters could be reduced. Then the residuals between the pre-prediction sequence and observation sequence could be corrected by the applying of the Fourier series. Last, after long-term adjustment based on the Markov state transition model, it proposed a combination method that could predict and correct the traffic flow data. This made the forecasted results not only reflecting real-time, but also reflecting the accuracy and adaptability. In the process of the instance validation, because of the characteristics of the traffic flow data such as randomness, self-organization and class cyclical, we must make denoising smooth for preprocessing over the actual measurement data before the implementation of forecasting. Experiments show that this model used both information of the current way to change the characteristics provided by real-time data and the overall law information of the way to change provided by historical data. The comparison result of model validation shows that the proposed model of the paper has better prediction accuracy.