Research On Urban Road Travel Time Prediction Method Based On Deep Learning

Accurate prediction of travel time can provide guidance for traffic management and travel decision.As a data mining method,deep learning has good applicability in forecasting.This paper aims to use the advantages of deep learning in data mining to predict the travel time of urban roads based on Chengdu floating car data.Firstly,based on the trajectory data of Chengdu ride hailing,the paper selected the main road of Chengdu Xinhua Avenue with a length of about 3km as the research object and the vector road network was built.After the data was cleaned,the paper filtered out the data according to the idea of constructing irregular regions by dividing the grid,transformed the coordinates and divided the data set according to the distribution of time and space interval.After that,fast matching was achieved by determining the candidate road segment set,the initial order matching,and the forward-backward check.On this basis,key information such as travel time was extracted to construct a sample set for deep learning.Then,based on the understanding of the existing deep learning methods,the paper concluded that the current deep learning prediction methods had some shortcomings,such as insufficient input and the structural explanatory was not enough.In order to improve these shortcomings,the paper analyzed the spatiotemporal characteristics of travel time,and concluded that there was a certain regularity in the time distribution of the travel time and a certain relationship between the travel time of the adjacent sections in the space.On this basis,the paper used sub-sections as the deep learning structural unit,optimized the model input based on time characteristics and the model structure by spatial characteristics,and designed a two-layer model architecture including traffic characteristic learning layer and travel time prediction layer.Then,according to the model architecture,the paper established a deep learning network model based on spatiotemporal characteristics.In the traffic characteristic learning layer,the road segment was the research object,including two steps: Firstly,the features such as the average travel time,the variance of the travel time,the average number of sampling points and the reliability were selected as the traffic characteristics indicators.Secondly,based on the time characteristics of travel time,the long-term historical and recent real-time feature inputs were determined,and then constructed the long-short-term feature fusion model by LSTM model and the multi-task learning idea in deep learning.In the travel time prediction layer,the bidirectional LSTM model was established based on the space characteristics of travel time,and the bidirectional time interval vector matching the model features was proposed.Then the double loss function was defined to provide supervised information for the model.Finally,the paper optimized the training process from gradient descent,learning factors and other aspects,and selected the best structural hyper-parameters through multiple training.The traffic feature learning layer was a LSTM network with 2 layer and 108 units,the travel time prediction layer was a bidirectional LSTM network with 1 layer and 108 units.The paper used the trained model,the common LSTM model and LSTM-DNN model to predict the data from 16:00-18:00 on the rest day and the working day,and evaluated the error indicators,the amplitude and trend of change,and overall distribution.The results show that the overall performance of our model was better than the other two models,which can better captured the characteristics of traffic changes and the applicability more stable.At the same time,it was found that there was a large error in the special situation when the frequency of travel time was relatively low,such as the particularly smooth time in the early morning or the particularly congested time in the peak hour.But even in special cases.The model performance of this paper was still better than the other two models,indicating that the LSTM model based on spatiotemporal characteristics proposed in this paper had certain advantages.