Research On Models Of Short-term Traffic Flow Forecasting For City Road

Short-term traffic flow forecasting is an important support for traffic guidance of city roads and plays a key role in the foundations of Intelligent Transportation System. Changes of the short-term traffic flow are influenced by various factors. Therefore, the use of traditional modeling approach is not only more complicated, but also the established model is difficult to migrate. The machine learning methods can make the model automatically find change rules in the data by training it with the historical traffic flow data set. So it can predict the traffic flow in the short term effectively.Short-term traffic flow forecasting models are studied in this paper, which is respectively based on the Support Vector Machine (SVM), Random Forest and Deep Neural Networks (DNN). The main contents and research results of this dissertation are as follows:(1) On the basis of the factors, that influence changes in the traffic flow, feature variables are extracted and a training set is built. Then a compromise data cleansing method is used base on the distribution of data integrity. The traffic flow data can be discarded in this method, when the observed rate is lower than 90%.(2) The traffic flow is predicted on the basis of SVM model. An improved random search algorithm is proposed to the problems of SVR/s hyper-parameters adjustment. It enhances the efficiency of hyper-parameters adjustment and reduces the complexity of the model use. Meanwhile, a higher forecasting level can be also achieved.(3) The random forest model is used to predict the traffic flow. It is adjusted with the default hyper-parameters and hyper-parameters optimization experiments. The results demonstrate that the random forest is ease of use in the traffic flow forecasting and easy to tune the hyper-parameters. It also has a higher forecasting accuracy and takes less operation time. Then the feature variables importance is estimated based on the random forest and a traffic flow forecasting method is proposed, which selects feature variables dynamically. So that the efficiency and adaptability of the model can be improved.(4)The DNN application in the traffic flow forecasting is discussed. The traffic flow data set is utilized to train the DNN respectively in two forms of layer-wise algorithm as well as the DNN containing ReLU directly. One of them gives the best forecasting results in this paper.