Traffic Speed Prediction Based On Spatio-Temporal Information Analysis And Machine Learning

With the rapid increase in the number of motor vehicles,traffic speed prediction has become more and more important.Accurate and reasonable prediction of traffic conditions can effectively alleviate traffic congestion,reduce waiting time,and assist managers in traffic management and public safety.At the same time,traffic speed prediction still faces many difficulties.The existing research lacks a reasonable and effective method of selecting spatio-temporal information,fails to fully explore the spatio-temporal characteristics of traffic speed,and can not fully express the differences in the contributions of different historical moments and different relevant road segments.In response to these problems,the main research contents and innovations of this thesis are as follows:(1)Considering that the spatio-temporal information is the basis of traffic prediction and the existing research lacks a method for selecting key information from mass information,the method of spatio-temporal region of support is proposed in this thesis.For the temporal region of support,a reasonable range of time intervals with strong correlations is determined by calculating the correlation coefficients of different time intervals and analyzing these correlation coefficients by cumulative distribution function;For the spatial region of support,fully considering the effects of delay and non-linear correlation,a sliding window correlation coefficient calculation method is proposed to search the relevant road segments of the target road.The analysis and construction of support region is universal and provides a basis for the prediction models.(2)Considering that the existing methods do not fully express the spatio-temporal correlation of traffic speed,especially the dynamic characteristics of traffic speed are not considered,this thesis proposes a spatio-temporal broad learning networks for single-step traffic speed prediction.This method first extracts the dynamic and static slow features of the variables in the spatio-temporal region of support,fully explores the changing trend and speed of traffic speed information;Then the method enhances the model's capabilities of nonlinear expression by random nonlinear enhancement nodes.On this basis,L1 regularization is introduced to sparse features and automatically select important features,and L2 regularization is introduced to prevent overfitting.This method has the characteristics of high prediction accuracy and short training time.(3)Considering that different historical moments and different related segments have different effects on prediction,this thesis proposes a dual-stage attention based sequence learning networks for multi-step traffic prediction.The method is based on encoder-decoder networks.The first-stage attention mechanism is introduced at the input of the encoder part to adaptively learn the contribution weights of different road segments at each moment,the attention at this stage is called spatial attention;For the decoder part,attention mechanism is introduced to adaptively learn the contribution weights of different historical moments to the current prediction moment,the second-stage attention is called temporal attention.Our method expresses the spatial-temporal correlation characteristics more finely and has higher prediction accuracy.And its prediction performance decays more slowly than other methods with the increase of the prediction time step.