Design And Implementation Of Travel Time Estimation Based On Neural Network

The project is based on the research project of my advisor Ruipeng Gao and the actual project during my internship in DiDiChuXing Company.With the application and development of intelligent transportation system,the travel time of the road has become one of the important reference indicators for people's travel choice.In the actual traffic application scenario,the travel time of the road is an important indicator to measure the road running state,traffic planning,intelligent traffic research and development,and urban traffic operation level.Accurately estimating the travel time of the road can not only improve people's travel efficiency,but also further improve the service quality of the traffic.The estimation method of road travel time mainly focuses on the origin-destination(OD)of the path.It is mainly divided into two categories,one is the path-based estimation method,and the other is based on the data-driven estimation method.Through a large number of experiments,the current method has relatively many limitations in terms of data sparsity,model generalization ability and data utilization.In order to effectively solve the above problems,based on the existing methods,this paper proposes a road travel time estimation method based on neural network.In the method,the analysis utilizes multi-source heterogeneous data features such as GPS trajectory data,road network data,and external attribute data such as weather,date,etc.,and combines different data features into the network model for training and prediction,and obtains a prediction value of the network model.In the experiment of this method,we first analyze and model the existing multi-source heterogeneous data.Secondly,different data processing methods are adopted for different data features.For example,in order to further effectively extract the spatial characteristics of the road network data,we use the DeepWalk algorithm for the road network data,aiming at retaining the characteristics of the road network topology adjacency relationship,the segment number of road is vectorized.For the external feature,since it belongs to the class value attribute,it cannot be directly input into the network model,so we use the Embedding method to represent it in a vectorized manner,which not only preserves the semantic similarity,but also it reduces the data dimension.Finally,based on the characteristics of different traffic data,combined with the Long Short-Term Memory(LSTM),the deep learning framework PyTorch is used to realize the model of the estimated road travel time method.In the training and testing process of the model,different model evaluation indicators are used,such as Mean Absolute Percentage Error(MAPE),Mean Absolute Error(MAE),and Mean Square Error(MSE)to evaluate the performance of the model.Through the training and testing of the datasets in Beijing and Shanghai,the MAPE values of the model are 0.1736 and 0.1807,the MAE values are 103.66 and 112.35,respectively,and the MSE values are 13279.21 and 13762.18,respectively.Through the experimental analysis of different existing methods,the road travel time prediction method proposed in this paper has lower prediction error,and is superior to other existing methods.