Urban Travel Pattern Mining And Application Based On Private Car Data

At present,some progress has been made in the mining and application of urban travel patterns,but the research based on private car data still has limitations.Existing studies on private cars mainly extract the travel data of some private cars,but there is no comprehensive study on full private car travel data.Different from partial private car data,full private car data can be used to discriminate,analyze and compare the characteristics of different private car travel patterns,providing effective decision support for urban management and upper-level planning.The existing research on travel patterns based on some private car travel data is not systematic,comprehensive and in-depth,and the relevant research methods still have limitations.In addition,private car travel data has important application value,and relevant studies have limitations in methods or scenarios,such as the matching of commuter private car carpooling and the prediction of private car traffic.Private car travel data are large-scale spatio-temporal data,so how to select and optimize the model to achieve the application purpose is worth studying.Therefore,this thesis systematically carries out research on private car data.In view of the current perception of private car mobility,we consider the following aspects:commuting private car travel is an important part of private car travel,and also an important cause of urban road congestion;At the same time,with the development of instant travel services,a large number of private cars have sprung up offering taxi-like services.Most instant travel services are still limited to one passenger per car,and the pressure on the roads is increasing.Therefore,it is necessary to identify and analyze them.The application of private car travel data: considering that carpooling among private car commuters can effectively reduce the flow of private cars on urban roads,thus relieving the pressure of urban road traffic;At the same time,traffic prediction is an indispensable part of urban management.It is a new idea to predict traffic flow by considering the components of different traffic flows,including different vehicle types and different travel purposes.The main innovations and contributions of this thesis include the following aspects:1.For the identification of private car commuters,hierarchical clustering method based on regular behavior is proposed.According to the diversity of private car commuters and the regularity of individual private car commuters,this method reasonably sets three constraints of private car commuters based on regular behavior.At the same time,a bottom-up hierarchical clustering algorithm is designed for the extraction of regular behaviors.The robustness of this method is enhanced by introducing adaptive threshold.The experiment shows that compared with the feature-based clustering method,the proposed method can detect private car commuters more accurately and comprehensively.2.Aiming at the problem of suspected operating vehicle identification,a suspected operating vehicle identification model(STL-Detector)based on self-taught learning is designed.The feature learning component of the model is designed as an autoencoder structure of Three Dimensional Convolutional Neural Networks(3-D CNN),and is trained using a large amount of unmarked auxiliary vehicle trajectory data.It can effectively learn the advanced feature representation of vehicle trajectory.The supervised classification component is designed as a random forest classifier,which is trained with a small amount of labeled data and can effectively classify suspected/non-suspected operating vehicles.Experimental results show that compared with the method without using unlabeled data,the recognition effect is significantly improved.3.A carpooling matching method based on reinforcement learning is designed to solve the carpooling matching problem of private car commuters.This method formalizes the carpooling matching problem of private car commuters into a Markov decision process.On this basis,a reinforcement learning process is designed,in which special processing is designed for initialization,action design and environment update.Its characteristics include: Using Double Deep Q Network(Double DQN)approximate the value function,so only a few parameters can be used to;Automatically assign driver roles;In the process of action design,time and space constraints are incorporated to ensure the rationality of the matching scheme.The experimental results show that compared with the existing combinatorial optimization solutions,it can effectively deal with the problem of carpooling matching.4.A combined prediction method(TSA-SL)is proposed for short-term traffic flow prediction.The method first decomposed the traffic flow into periodic part and volatility part,and then predicted the future periodic value and volatility value respectively.In the periodicity part,we use Fourier transform(FT)to model periodicity.In the volatility part,SVR,GBRT and LSTM are used to model the volatility.FT-SVR,FT-GBRT and FT-LSTM combined prediction models are established according to TSA-SL method.The experiment shows that the combined prediction method is superior to the single prediction method in predicting short-term overall traffic flow,private car traffic flow and commuter traffic flow.