Traffic Flow Probabilistic Prediction Using Deep Temporal Convolutional Networks And Copula Model

Traffic flow prediction,as a key technology in the intelligent transportation system,not only provide travelers with real-timebus arrival time,traffic speed,and other traffic information services,but also provide technical support for signal control and path guidance in road traffic management by assessing road network status.With the dynamic openness of the urban traffic system,the difficulty of predicting traffic flow accurately is increasing as it’s great complexity,randomness,and uncertainty.As a result,a more plausible prediction method is necessary for describing and predicting the randomness of traffic flow parameters to improve the road capacity and traffic stability.Deep neural networks,a new technique,can reveal intrinsic laws and characteristics from multi-source traffic flow data with powerful data learning capacity.However,their deterministic prediction result can hardly capture the randomness of traffic flow.To overcome this problem,this thesis proposes a deterministic prediction model of traffic flow based on Deep Temporal Convolutional Network（DeepTCN）.On this basis,with the idea of ensemble learning Staking,the DeepTCN model and the Copula model are integrated based on the ensemble strategy of probabilistic error compensation using.Thus,the probabilistic prediction results can be obtained by calculation through the fusion model,which also improves the prediction accuracy of the DeepTCN model to some extent.The main research contents are given as follows.First,data analysis and processing are performed based on the real traffic big data collected from Guangzhou.On the one hand,the spatiotemporal characteristics of the traffic flow parameters of the road segment are examined by using multi-level processing techniques such as association and combination of highway detector data,outlier processing,and standardization.Then,the flow-density relationship curve can be obtained via polynomial curve fitting.Finally,the probability distribution functions of velocity and flow under various densities are acquired using the kernel density estimation approach.On the other hand,bus headways are determined by cleaning the data of buses’ entering and exiting a station,and the calculation of traffic flow characteristics.By the analysis of bus states and the spatiotemporal features of bus headways,the existence of randomness in the bus operating process is proved through an approach of kernel density estimation.Second,a deterministic traffic flow prediction model based on DeepTCN is established using the theory of deep learning.A residual neural network based on stacked dilated causal convolution is introduced in the model,which can capture the correlation of long time series.The applicability of DeepTCN is illustrated in two scenarios（prediction of traffic speed on expressway and prediction of headway of bus）,where their total prediction performance for time series tasks outperforms classical neural network models.Finally,a probabilistic prediction approach based on Copula theory is proposed for prediction error compensation to address the problem that the deterministic prediction outputs of temporal convolutional networks cannot reflect the randomness of traffic flows.The joint probability distribution of the predicted and observed variables of the DeepTCN model is constructed based on the correlation between the two variables.The conditional probability distribution function of the prediction error under the conditions of the predicted variables is fitted to compensate for the error probability and obtain the probability prediction result.It was showing that the proposed fusion model can not only reflect the randomness of traffic flows but also give more decision-making information for engineering applications.