Research On Taxi Demands Prediction Based On Deep Learning

With the continuous growth of urban scale and population,urban traffic problems are becoming increasingly prominent.Traffic congestion,difficult to take a taxi and other problems make the travel time cost higher and higher.As a common means of transportation in residents’ daily travel,taxi plays an important role in urban transportation.It is of great significance to realize the accurate prediction of taxi demand for taxi dispatching management,reducing the empty driving rate,saving energy consumption and maximizing the satisfaction of passenger travel demand.Nowadays,taxis are generally equipped with GPS equipment,and a large number of GPS trajectory data will be generated during the driving process of taxis,which makes it possible to predict taxi demand based on deep learning method.In this thesis,two taxi demand forecasting methods based on deep neural network are proposed,the main research contents are as follows:(1)A taxi demand forecasting method based on multi-task learning and graph convolution network(GCN)is proposed.The method constructs the graph based on two kinds of spatial relationships between road segments,including the local relationship between adjacent road segments and the global relationship between long-distance road segments with similar regional functions in the urban road network.The local relationship graph of departure flow and the global relationship graph of departure flow are constructed correspondingly.And the spatial characteristics are realized based on Graph Convolution Network(GCN)and Gated Linear Units(GLU).Based on Long Short-Term Memory(LSTM),time feature mining is realized.In addition,multi-task learning strategy is introduced,which takes the prediction of taxi arrival flow as the related task,constructing the local relationship graph and the global relationship graph of taxi arrival flow,so as to help reduce the error of taxi departure flow prediction and realize the prediction of taxi arrival flow.(2)A multi-factor taxi demand forecasting method based on GCN and TCN is proposed.In this method,three kinds of spatial relationship graphs are constructed,which are helpful to the full mining of spatial characteristics,namely,the local relationship graph of departure flow,the global relationship graph of departure flow and the relationship graph with origination-destination(OD)flow of departure flow.Through the construction of relationship graph with OD flow,the network can find the specific OD flow relationship between road sections caused by residents’ travel behavior.The network realizes spatial feature mining based on GCN and temporal feature mining based on Temporal Convolutional Network(TCN),which avoids the difficulty of LSTM training.In addition,the method considers the influence of weather,day of week,air quality,temperature on taxi departure flow,and uses Fully Connected Layer to extract its features.In this thesis,the taxi trajectory data of Xi’an city is used to verify proposed methods.The experiment results show that the two taxi demand forecasting methods proposed in this thesis are better than the common traffic forecasting models.