Study On Short-term Speed Prediction Of Expressway Based On Spatiotemporal Correlation

With the development of intelligent transportation system,traffic monitoring equipment on urban road is widely popularized.Vehicle-mounted navigation equipment and positioning system based on mobile terminal are basically covered.These facilities provide great convenience for collecting traffic data.Based on the real-time data,accurate traffic speed prediction is of great significance for traffic control and guidance.It helps alleviate the uneven distribution of road spatiotemporal resources.Based on previous studies,scholars has made great progress on short-term speed prediction method,but there are also some problems including the lack of basis for the spatiotemporal dimensions selection,the poor adaptability of prediction model to different traffic conditions,and the error accumulation of multi-step prediction.In order to solve the problems,some new idea had been proposed in this paper including dynamic spatiotemporal dimension selection,adaptive construction of similar state training set and error correction in multi-step prediction.The research contents were as follows.(1)Two types of average speed collecting methods were introduced in the paper,and we proposed the processing methods for information matching,data screening,and data repair.According to the processed time series,the spatiotemporal distribution characteristics of speed were analyzed.The results showed that there was a certain similarity in the speed of adjacent road sections in the spatial dimension.In terms of the time dimension,the change of traffic status in the downstream lagged behind that in the upstream.(2)According to the spatiotemporal characteristics of speed,we proposed a dynamic selection method of the spatiotemporal dimension for training set.Because the road environment and traffic status of each target section were different,we determined adaptive matrix dimension for each target section.The equivalent distance was used to measure the influence of spatial neighborhood.Cross-correlation was used to determine the lag degree of time window of the surrounding section and the AIC during the ARMA modeling process was used to determine the lag degree of the target section.The spatiotemporal matrix was adaptively constructed.(3)A short-term speed prediction model considering similar traffic features was proposed.The model included three parts: similar traffic features extraction based on Gaussian weighted distance,short-term speed single-step prediction based on PSO-SVMGARCH,and a multi-step speed prediction model based on historical data correction.The purpose of similar feature extraction was to determine a suitable training set for the prediction model.PSO algorithm and GARCH model were used to improve the efficiency and accuracy of SVM algorithm.To solve the problem of error accumulation in multi-step prediction,we used the probability distribution of historical data to control the prediction results.(4)The model was validated by the measured data on Changchun West Expressway and Seattle Highway 167.We compared the prediction accuracy among different combinations and different algorithms.The results showed that the PSO-SVM-GARCH model with dynamic selection of spatiotemporal dimensions had better prediction performance.With the correction of historical data,the cumulative error of multi-step prediction was also significantly improved.