Research On Traffic Flow Prediction Methods Based On Deep Spatio-temporal Models

The task of traffic flow prediction is to estimate the averaged number of vehicles in a specific region and a future time interval given the historical flow data from the global traffic network.Traffic flow prediction is a fundamental component in intelligent transportation systems.The prediction accuracy is a significant part of Advanced Traffic Management Systems(ATMSs).However,many existing prediction models endure several shortages,although the research on it has been going on for decades.Most of the methods are constructed as a shallow model,which is difficult to reveal the intrinsic spatio-temporal relations embedded in traffic raw data.Moreover,the separation of feature learning and predictor learning brings a sacrifice of model performance.Then the hand designed features are difficult to be tuned appropriately.Finally,few existing methods consider the incomplete data problem which is in fact very severe for practical application.In this thesis,we research traffic flow prediction based on the deep learning methods,aimed at the shortages mentioned above proposed,for the purpose of improving the traffic flow prediction accuracy.Our work can be summarized as follows.At the beginning,we summarize the current situation of traffic flow prediction at home and abroad,and find out four shortages of existing traffic flow forecasting methods.Then,we analyze the characteristics of traffic flow data especially the spatio-temporal correlation,and study on the methods of traffic data preprocessing.Furthermore,we design two traffic prediction methods based on deep learning,which are applied random subspace and sparse representation to solve the incomplete data problem and explore the spatio-temporal correlation.In addition,we improve the hyperparameter random search algorithm used in the convolution neural network.The improved algorithm use uniform design to achieve the best effect of the search algorithm,which makes the best effect of the model training.Finally,a large range of experiments with various traffic conditions have been performed on the traffic data originated from the California Freeway Performance Measurement System(PeMS).The experimental results corroborate the effectiveness of the proposed approach compared with the state of the art.The contributions of this work are summarized as the four points.First,we transform the time series analysis problem into the task of image-like analysis.In addition,we design an ensemble learning strategy via random subspace learning to make the model be able to tolerate incomplete data.Benefitting from the image-like data form,we can jointly explore spatiotemporal relations simultaneously by the two-dimension convolution operator.Then,we propose the spatio-temporal correlation measurement method based on sparse representation,and construct sparse similarity matrix according to it.Furthermore,we propose an improved random search based on uniform design in order to optimize hyper-parameters for deep Convolutional Neural Networks(deep CNN).Finally,we propose a correlation similarity metric learning method based on graph embedding to measure the similarity between observation points and expand the kernel direction.