Research On Taxi Demand Forecast Based On Temporal And Spatial Data Mining

With the improvement of the urban economy and people's living standards,taxi has become a travel choice for more and more families due to its convenience and comfort.However,the increasing demand and quantity of taxis will lead to the imbalance between supply and demand,traffic congestion,energy waste,and other problems.Therefore,how to accurately predict taxi demand has become a research hotspot in intelligent transportation.The prediction of taxi demand cannot be separated from taxi trajectory data,which belongs to the category of spatiotemporal data and has a complex spatiotemporal dependency,making it difficult for traditional methods to effectively model.This thesis starts from analyzing multi-factor affecting the future taxi demand and modeling of spatiotemporal dependency,and by means of the powerful highly nonlinear spatiotemporal dependency modeling ability of deep learning methods,two prediction models based on the deep learning methods are proposed.The main research contents and results are as follows:(1)A deep bidirectional spatiotemporal network model based on spatiotemporal data mining is proposed.This model first builds multi-layer local Convolutional Neural Network(CNN)and gated CNN to capture the spatial static and dynamic dependency,and then utilizes deep bidirectional LSTM to learn forward and backward temporal dependency.Besides,this model extract short-term and long-term traffic volume sequence to model periodicity.(2)A multi-factor spatiotemporal graph CNN based on spatiotemporal data mining is proposed.On the one hand,this model extract history inflow sequence and history demand sequence from the taxi trajectory to model multiple temporal dependencies.On the other hand,we establish geographic adjacency matrix and potential dependency matrix between regions to model multiple spatial dependencies.Furthermore,based on the spatiotemporal convolutional blocks,this model constructs adjacent sequence component,daily sequence component,regional potential correlation component,and region inflow component to capture the potential spatiotemporal representations of different factors.Finally,after simulation experiments on multiple public datasets,it has been proved that the two models proposed in this thesis have lower prediction errors than the existing methods under several evaluation metrics such as Root Mean Squared Error(RMSE),Mean Absolute Percentage Error(MAPE)and so on,which shows that our models can better solve the problem of taxi demand prediction.Moreover,our models can easily extend to other spatiotemporal data prediction tasks.