Bus-to-station Time Prediction Based On Multi-source Bus Data And LSTM

In the current domestic bus operation system generally provides only the first and last station departure time environment,accurate prediction of bus arrival time is very important for passengers who want to determine the departure time and reduce the waiting time.At the same time,it can improve the efficiency of the bus scheduling system effectively.In order to further improve the medium to long time bus station and the accuracy and reliability of the existing prediction methods and basis,using the bus multi-source data,set up a method of based on the depth of the LSTM learning for the multi-step prediction of bus travel time,and this method on the measured data of calibration and verification.Firstly,based on the operation mechanism of bus-to-station problem,this paper establishes an arrival time prediction model based on multi-source bus data,and defines the arrival problem as a multi-step prediction problem of time series.According to the characteristics of the multi-step prediction problem and the arrival time prediction model,the LSTM long and short time memory algorithm is selected,and the recurrent neural network is taken as the core to predict the arrival time of the bus from the perspective of learning the long-term dependence of traffic.Secondly,this paper studies the existing multi-source bus dataset.Given that the current dataset only contains static data and historical GPS operational data,it has limited help in studying bus operation.This paper extracts the corresponding dynamic factors from the multi-source bus information.The extracted dynamic factors are merged into the vector space,which ultimately forms the training set used in this paper.Taking the actual data as an example,the LSTM(Long Short-Term Memory)prediction model proposed in this paper and the other three mainstream prediction models(BP neural network,SVR,KNN)are in five static influencing factors(operation period,weather conditions,road infrastructure,intersections and peaks).On the basis of NAND),two dynamic influencing factors related to bus operation status(number of people on the station,bus card type)were designed,and the influence of multi-source bus data interaction on bus arrival time was analyzed..Finally,taking the actual data as an example,LSTM model is compared with the resultsof BP neural network,SVR,KNN and other prediction models according to the data input and output schemes of different dimensions,and the model performance is analyzed and improved.The results of comparison with other traditional models show that,different from the prediction idea based on single data or regression,LSTM can capture the long-term dependence of traffic hidden in the sequence of bus operation and learn the relationship between the changes of various factors and bus operation.The extension of the network corresponds to the operation of the bus.When new elements are input,the model can correct the prediction in the previous step and has strong dynamic adaptability.Experimental comparison with common algorithms also proves the effectiveness of the multi-step prediction algorithm based on multi-source bus data.