Traffic Flow Prediction Based On Deep Gaussian Process

As a Deep Learning methods proposed in recent years,the Deep Gaussian Process(DGP)is widely used in the unsupervised learning field due to its powerful learning ability to represent the implicit details of nonlinear data.Different from traditional Machine Learning methods or Deep Learning methods,starting from a single layer Gaussian Process,the DGP can solve the difficulties of analytical intractable solution facing compute the maximum posterior probability distribution by inferring the approximate solution of posterior distribution through variational inference.Thus,besides giving the accuracy predictions,it can also represent the uncertainty of predictions in probabilistic.This characteristic shows significant advantage in the prediction of time series,especially in the prediction of traffic flow.The main work in this thesis as follows:(1)For high-noise characteristics of data,large amount of works in data preprocessing are performed in this thesis.These works include:by comparing the different filling methods,the Forward Filling method is selected.For the high noise in the data,several commonly used filtering algorithms are compared and the local linear weighted regression filters is selected to de-noise data.(2)A three-layer DGP network composed of different kernel functions is built to predict the long-term traffic flow in real dataset and reaches a satisfactory result.(3)For the superiority of the DGP,comparing the DGP with the single layer Gaussian Process,the DGP shows better characteristics in both prediction accuracy and fitting the nonlinear feature in data.Especially,when facing the change of traffic flow,the DGP can learn the details in data while the traditional Gaussian Process cannot.(4)The prediction results between the DGP and the classical Deep Learning network LSTM are compared laterally in this thesis.The experiment shows that both network have a strong ability to represent the details of the data in-sample.Out of the sample,the DGP is more robust because it has the advantages of the sparse parameters due to its network structure,while the overfitting phenomenon occurs in the LSTM.As a machine learning method with probabilistic background,the DGP is closer to the nature of human logical reasoning than the single layer Gaussian Process.In realistic predictions,it shows that the DGP can learn the non-stationary and non-linear features of multiple superposition in traffic flow data.While achieving good accuracy,it can also infer the uncertainty confidence interval of the prediction.Thus,in the field of time series,especially in the traffic flow prediction,it shows greater advantages than the traditional singer layer Gaussian Process.