Research On Prediction Of Traffic Flow Based On Dynamic-fuzzy Neural Networks

Traffic flow prediction is an important field within intelligent transportation system.When accurate and real-time, traffic flow prediction serves as an important prerequisite takingcontrol over traffic signal. Besides, it is also conducive to inducing traffic flows in this case.Therefore, how to improve traffic flow prediction methods and improve its accuracy hasbecome a hotspot research.It is known that computation intelligent methods like neural network and fuzzy systemhave been widely used in traffic flow prediction and by which better prediction effects havebeen gained. By reference to the advantages of both neural networks and fuzzy system, thethesis studies the traffic flow prediction under the guidance of Dynamic-Fuzzy NeuralNetworks (D-FNN) method. Furthermore, based on chaos theory, the thesis makes acomparative analysis between neural network method and Adaptive Neural Fuzzy InferenceSystem (ANFIS) method in terms of prediction performance by applying D-FNN method totraffic flow time series and video network flow time series. The research results show thatD-FNN method is effective.The present thesis is composed of the following aspects:(1) Based on the neural networks and fuzzy logic theory, the thesis studies ANFISmethod; based on the previous studies, the thesis takes a further look at D-FNN method andits learning algorithm(2) By introducing chaos theory to traffic flow prediction, the thesis studies chaoticphase space reconstruction theory. By reference to the principle that the dynamiccharacteristics of traffic flow are predictable and analyzable, the maximum Lyapunov index iscalculated in order to determine whether traffic flow’s time series is of chaotic characteristicor not; the thesis also focuses on the embedded dimension of the time series and the selectionof delay time parameter. The Cao method and mutual information method are used todetermine the embedded dimension and the delay time respectively.(3) To begin with, the thesis discusses D-FNN method’s performance in multi-stepprediction of Mackey-Glass chaotic time series. A multi-step predictive model is establishedby using the and ANFIS method. With regard to the multi-step prediction, a comparativeanalysis is then conducted. Next, according to the short-time traffic flow released by amonitoring station of Beijing and Britain’s transport network and the video stream byWurzburg University of Germany, the predicting experiments are made by applying D-FNNpredicting model to these three groups of time series and meanwhile predictive effects areanalyzed. As far as the predicting performance is concerned, a comparative study is madebetween the predicting experiment above and the predicting model built by Back-Propagation Network and ANFIS method. Lastly, the thesis studies the predicting performances of thesepredicting models mentioned above as embedding dimension and delay time are different.The study results show that the predicting model based on D-FNN method is effective inpredicting chaotic time series and that when the predicting model of D-FNN and ANFIS arebuilt with optimal embedding dimension and delay time, their prediction accuracy is betterthan that of the predicting model of Back-Propagation Network. To sum up, D-FNN methodis found effective in predicting traffic flow time series and video streaming time series.