Research On Dynamic Path Guidance Method Of Urban Individual Travel Based On Deep Learning

With the increase in the number of motor vehicles in my country,the urban road network tends to be saturated.So that the per capita travel cost increases,the effective utilization rate of road resources decreases.And the public security traffic management department has a heavy task.Under the current traffic situation,there is an urgent need for research.That is,how to enable drivers to obtain timely and accurate traffic travel service information so that they can quickly adjust their travel routes with changes in traffic conditions.The optimal path algorithm adopted by the traditional intelligent transportation system usually does not consider the real-time traffic changes.Its overall control ability in time and space is limited,and it can play a small role in the current urban road network with various changes in traffic operation status.Recent years,the neural network model based on deep learning theory has strong time-space control ability,and has been widely used in the field of transportation.The path induction problem is studied in this thesis based on such methods.Firstly,the domestic and foreign research status of the research content is reviewed.The theoretical research on the time series forecasting model and path planning method based on the deep learning method is carried out.The LSTM model is determined as the basic prediction model of this thesis.On this basis,a bidirectional mechanism and a one-dimensional CNN network layer are introduced.And a combined CNN-Bi LSTM model is proposed.After adjusting the hyperparameters and network structure,the result errors MAE,RMSE and MAPE of the optimized model were reduced by 1.7%,7.7% and 3.7% respectively.At the same time,the combined model is trained with input data of 5min,10 min and 15 min granularity respectively.The influence of different time granularity on the prediction ability of the model is analysed.The experimental results show that: compared with the 5min particle size,the MAPE increases by1.32% under the 10 min particle size,and increases by 3.17% under the 15 min particle size.And compared with 10 min granularity,MAPE increases by 1.85% at 15 min granularity.It shows that the prediction error of the model increases with the increase of the granularity of the input data.Secondly,with the help of Arc GIS geographic information system and geographic information data,feature class setting,feature set construction,topological relationship processing,connectivity adjustment,attribute and direction setting are all completed.Thus,a path analysis network dataset is established.Visualization of path analysis is achieved.And the dataset can be used for subsequent analysis.Thirdly,the principles of several commonly used time series prediction models and their training and prediction processes are analysed in detail.ARIMA model,RF model,RBF neural network model,LSTM model and GRU model are used as comparison models.The speed data of the main line of Bao’an Avenue in Shenzhen in November 2020 is taken from Baidu trajectory data.After preprocessing in MATLAB,it is used to design training and prediction experimental models.The experimental results show that the prediction ability of the CNN-Bi LSTM combined model is the best.Compared with the comparison model,its errors RMSE,MAE,and MAPE are reduced by 33.2%,40.7%,and 35.5% on average.Fourthly,the route from Shenzhen West Railway Station to Bao’an International Airport is taken as an example.With the help of map data,Dijkstra algorithm is used to calculate the shortest path.Then with the help of the speed data of relevant road sections in November 2020 in Baidu trajectory data,the vehicle speed results predicted by the CNN-Bi LSTM combined model on November 30 th were combined with the map data to calculate the dynamic itinerary.It is introduced into the path cost calculation process of the traditional path planning method A*algorithm.And then the shortest travel time is used as the induction target,the dynamic prediction induction path calculation is performed.Based on three different time periods of morning peak,evening peak and peaceful peak,the dynamic prediction induced path results were calculated respectively.Their travel time was compared to the travel time of the shortest route.The experimental results show that: in the morning and evening peak hours,the travel time of the dynamically predicted route is shortened by 12% and 6.4% respectively.The reliability of the dynamic A* algorithm to calculate the dynamically induced route has been verified.Finally,the requirement for the induced response speed,and the tolerance of the travel time calculation to small differences in vehicle speed data are taken into account.The data granularity of the dynamic A* algorithm input is adjusted.The dynamic induction paths in each period were calculated with time windows of 5 min,10 min and 15 min,respectively.The experimental results show that the path results under different time windows are consistent.The average travel time difference between the 10 min time window and the 5min time window is 2.8%,and between15 min and 5min and 10 min is 4.2% and 1.4%.The robustness of the dynamic A* algorithm in the steady traffic flow is confirmed.The research results of this thesis can provide a certain theoretical basis for personalized travel services in traffic management.