A Missing Traffic Flow Recovery Method Based On Deep-learning

A steep increase in traffic data is observed with the rapid development of traffic informatization.The traffic flow(i.e.,number of vehicles passing a reference point per unit of time)plays a key role in traffic stream description,and the growth and decrease reflect basic properties of the traffic stream.However,inevitable and ubiquitous problems with missing data occur due to detector faults,transmission distortion,or storage errors.Missing data problem not only reduces the quality and validity of the raw data collected but also hinders the utility of subsequent data analytics performed on the data.The data imputation refers to the recovery process of substituting missing data values using statistical models and algorithms.On the one hand,it helps deepen the understanding of spatialtemporal correlation and statistical characteristics of traffic flow.On the other hand,it is of great significance to follow-up analysis of intelligent transportation system(ITS).Hence,researching on missing traffic flow data imputation has important value on theoretical and practical research.From the perspective of complex urban traffic network,this article combines the main characteristics of large-scale data,local spatial-temporal correlation and data dynamic missing,aiming at the random missing on spatial-temporal and long-term missing.The main research work and contributions of this article are as follows:1.In the aspect of local spatial-temporal correlation in road network,a local road network flow calculation was proposed.First,a data-driven road network matrixing algorithm is proposed to mine location relations from trajectory data and reconstructs a road network into a matrix.In the matrix,the correlation between locations is positively related to distance.Second,convolutional local calculation in convolutional neural network is introduced to calculate local road network flow,thereby reducing the computational complexity and improving the generalization ability.2.In the aspect of data dynamic missing,a traffic flow recursive data imputation neural network(FRDINN)was designed.Initially,this deep neural network method combines the temporal relevance of historical flow with spatial correlation of current fragmented flow data with two sub-network and one merging network.Second,to obtain more prior knowledge on temporal relevance,recovery results are spliced together to form a three-dimensional tensor,and fed into the neural network for recursive calculation to reach missing data recovery.At last,the thesis verifies the proposed model and algorithm through using real vehicle passage records collected from the traffic surveillance system of a provincial capital city.Experimental results demonstrate that the optimized model cannot only effectively utilize local spatial-temporal correlations of road network flow to improve the generalization ability,but also effectively combine and utilize spatial and temporal correlations to impute missing flow data.