Research On Multi-lane Short-term Traffic Speed Prediction Based On The Clockwork Recurrent Neural Network

Currently,traffic congestion has become a common threat in cities.The traffic pollution,traffic safety and other problems induced by traffic congestion deteriorate the development of cities.The intelligent transportation system(ITS)which can alleviate traffic congestion effectively has attracted more and more attention in recent years.As an important function of the ITS,traffic control and guidance can adjust the traffic management and control scheme in a micro sense,which depend on the results of short-term traffic flow prediction.According to the characteristics of expressway traffic flow,a multi-lane short-term traffic speed prediction model based on the CWRNN was proposed and gradually improved.This work which could balance the accuracy and efficiency of prediction contributes to expressway management and control.Firstly,this paper sorts out the research status of short-term traffic flow prediction,introduces the experimental sample data and analyzes its traffic flow characteristics.For one thing,the traffic flow data have experienced a simple integration to analyze the basic characteristics of three parameters;for another,the Pearson coefficient was used to analyze its correlation from two dimensions of time and space.Analyzed results can be used for subsequent data repair and input set construction of the prediction modeling process.Based on the characteristics of traffic flow,the prediction modeling should follow the principles of accuracy,efficiency and flexibility.Secondly,the traffic flow data preprocessing process is described.By dividing the "dirty data" in the original data into three categories: data missing,data redundancy and data error,the data are eliminated by using the standard time stamp and the identification method based on the traffic flow theory.Aiming at the missing data,a lane-level data repair method based on spatiotemporal RBFNN is proposed.The speed data of the Beijing Third Ring Expressway are taken as an example to verify the repair effect of the method.Thirdly,to the best of our knowledge,the clockwork recurrent neural network is introduced for multi-lane short-term traffic speed prediction for the first time to give consideration to the prediction accuracy and efficiency.At the same time,the random forest theory is used to calculate the importance of the input variables to determine the optimal lookback time window length,so a multi-lane short-term traffic speed prediction model based on the CWRNN is constructed.The results show that the accuracy,stability and efficiency of the model are better than other benchmark models in most scenarios.Finally,to further improve the accuracy of the model in multi-step prediction,the idea of sequence to sequence is introduced,so a short-term traffic speed multi-step prediction model based on sequence to sequence is constructed.Here,the data of the Third Ring Expressway are used to verify the prediction effect of the model when input different variables.The results show that in the case of single input variable,the introduction of sequence to the sequence framework can effectively improve the accuracy of multi-step prediction;in the case of multiple input variables,the addition of flow data can make the input variables contain more comprehensive traffic information,and continue to increase the attention mechanism can make the model better deal with the complex problems of multiple input variables,thus improving the accuracy.At the same time,the experimental results also confirm the effectiveness of introducing the idea of sequence-to-sequence to improve the accuracy of short-term traffic flow multi-step prediction.There are 51 figures,14 tables and 77 references.