Research On Traffic Pattern Recognition And Prediction Based On Mobile Phone Signaling Data

Traffic big data is the key technology for building an urban intelligent transportation system,driving the overall upgrade and innovative development of urban transportation in terms of perception,human-vehicle-road coordination,big data analysis,and comprehensive services.Signaling data has the advantages of timeliness,continuity and wide coverage.Through the research on traffic pattern recognition and prediction of signaling data,it is possible to analyze the flow of people and road network conditions,which has theoretical significance and practical value for the implementation of intelligent transportation big data services and the establishment of urban governance systems.At present,many studies have been carried out on signaling data at home and abroad.On the one hand,most of the travel pattern recognition methods based on machine learning do not consider the impact of the combined features of different inputs on the model performance.On the other hand,the traditional traffic flow prediction model does not comprehensively consider the impact of the temporal and spatial attributes of traffic flow,intraday time series and environmental factors on the performance of the model,which restricts the prediction accuracy of traffic flow.In view of the above problems,the main work and main contributions of this research are divided into the following three aspects:(1)Stop points recongnition based on signaling data.Firstly,the space-time travel chain is constructed based on individual signaling data.Secondly,by calling the map GIS engine to map the signaling data to the traffic cells under study,the space-time grid clustering algorithm proposed in this research is used to constrain the user’s parking behavior in space and time,and a user stop point recognition model based on space-time clustering is formed.Finally,identify the stop points in the user’s travel trajectory,and complete the division of the travel chain.The simulation results show that our method is suitable for signaling data with unstable sampling interval and poor positioning accuracy,and the accuracy is improved by 2.9% compared with the ST＿DBSCAN algorithm.(2)User travel pattern recognition based on signaling data.Firstly,divide the reconstructed user travel trajectory,and extract travel trajectory features such as average speed,acceleration,and navigation data.Secondly,extract membership function and clustering features of user travel trajectory based on membership function and clustering algorithm.Finally,based on the above features,a multi-dimensional feature set of user trajectories is established,and the XGBoost model is used to evaluate the impact of different feature combinations on travel mode recognition in different scenarios.The simulation results show that the accuracy rates of traditional combined features and undifferentiated periods are 83.1% and 87.3%,and the accuracy of our method is about 91.75% in peak hours and 90.35% in off-peak hours.The performance of our method is better.(3)Traffic flow forecast based on signaling data.Firstly,a feature matrix is constructed based on the traffic flow data combined with the weather conditions of the road.Secondly,combining the advantages of CNN spatial expansion and LSTM long-term memory,use CNN to extract spatial features in traffic flow data,and use the output spatial features as the input of LSTM to extract temporal features in traffic flow data.Then,the spatiotemporal characteristics of the traffic flow are used as the input of the regression prediction layer to calculate the prediction result corresponding to the current input.Finally,an attention mechanism is introduced on the LSTM side to make the model focus on learning more important data features and further improve the model prediction accuracy.The simulation results show that the prediction results of the CNN-LSTM-Attention model considering weather factors are better than those without considering weather factors.At the same time,compared with the traditional neural network model,taking the sampling interval of 5 minutes as an example,the MAPE value of this model is about 7.63% lower than that of CNN(12.06%),LSTM(10.61%)and CNN-LSTM(8.17%).