Dynamic OD Prediction For Urban Networks Based On Automatic Number Plate Recognition Data

OD flow reflects the traffic distribution between the Origin and Destination in a certain time interval.OD traffic flow information is essential for effective traffic planning,operation,management and control.Among them,short-term(predicted)OD flow is an important piece of information for accurate and effective dynamic traffic management as well as control measures.Dynamic OD traffic forecasting pays attention to the trend of OD matrix in time series,and knowing this trend is a necessary condition for realizing real-time traffic management and information system.Based on the different definitions of the randomness or chaos of traffic flow information,OD traffic forecasting methods which have been proposed in the literature mainly include parametric and non-parametric approaches.These two methods focus on the periodic change and short-term fluctuation of OD traffic data respectively,and have different emphasis on prediction accuracy and efficiency.Most of the parametric methods are based on a set of state-space models,while non-parametric models often use pattern recognition and machine learning methods to achieve short-term prediction.Firstly,the high-dimensional characteristics of OD matrix in urban network are analyzed in this paper,and a data reduction method based on principal component analysis(PCA)is proposed.Five principal components can represent more than 83% of the original OD data with the dimension of 462.Then,based on the reduced dimension principal component data,the structure pattern and deviation are obtained by the polynomial fitting approach.The recursive state space model is established,and the prediction model(parametric)is proposed based on the Kalman filter method.In addition,three prediction algorithms based on K-Nearest Neighbors are proposed in this paper.The first prediction method based on KNN uses the principal component data directly,and takes the difference of the principal component scores between adjacent time periods as the trend of the pattern to construct the state set and label set.The second and third methods use polynomial fitting to derive the state pattern set and label set data and obtain the structure model set and the residual set.The difference between the latter two K-NN methods is the prediction of the variance of residual.The second K-NN model assumes that the residual is Gaussian white noise,and the variance of the residual in the latter period is the same as that in the previous period.The third K-NN model predicts the variance of the residual by applying the K-NN method again.The proposed four approaches are validated with three days' field ANPR data from Changsha city,P.R.China.The results show that on one hand our proposed approaches are able to make accurate prediction of OD flows under different demand conditions.On the other hand,the prediction accuracy is highly dependent on the quality of the available OD data: the Kalman filter model performs better for regular and periodic OD patterns;while for irregular OD matrices K-NN models could make more accurate prediction.