The Application Of Deep Learning On Short-term Traffic Flow Prediction

Accurate and real-time traffic flow forecasting has important practical significance in coping with the traffic congestion and traffic jam,meeting the demand of human beings for efficient travel,and also in problems of efficient traffic control and driving route guidance.In this paper a study is conducted on employing the Deep Learning technology into short-term traffic flow forecasting.A series of experiments are done to compare the various deep learning based models to show their applicability and advantages in short-term traffic flow prediction and then a forecasting model which has composed multiple deep learning networks is proposed.In this paper,the main works are as follow:Firstly,applying the deep learning method into short-term traffic flow prediction is done by way of predicting traffic volume with several kinds of deep learning based models.It is turned out that the deep learning methods for traffic flow forecasting is feasibility and effectiveness.Analyzing the advantages and existing problems of deep learning approach in short-term traffic flow prediction problem is done meanwhile flowing by a deep analyze of one of the experimented models who performs badly.Then,put forward a method for short-term traffic flow prediction.The method is given in form of combining several predicted values given by different kinds of deep learning based models and here the combining is accomplished by using the theory of bays.Taking the significance of parameters in short-term traffic flow forecasting into consideration,the degree of confidence of each submodel is estimated with the Bayes theory.And then this kind of confidence will be used when calculating a weighted mean among all the prediction values to be the final prediction value.Then it means a Bayes based Composed-Deep Learning-Predictor is put forward in this way.As shown in a series of comparison experiments,the BCDLP model has kept various advantages of its submodel so that it can improve the accuracy of a prediction task and ensure its performance of fitting at the same time.