Research And Implementation Of Urban Travel Demand Forecasting Based On Deep Learning

Reliable regional travel demand forecasting can provide reasonable and effective suggestions for the scheduling and planning of traffic resources.However,travel forecasting is a very challenging problem,facing massive spatial-temporal big data modeling problem.And how to extract the spatial and temporal features of the data effectively has become a research hotspot of urban computing.Therefore,this thesis makes a systematic study of travel demand in urban areas by using online car-hailing order data and combining with deep learning methods.The main work of this thesis includes the following three aspects:1.The order data of online car-hailing is processed and spatial-temporal characteristics are analyzed.First,the order data is deduplicated and divided spatial-temporally to obtain the travel time-space data used for regional travel demand forecasting.Then the spatial-temporal characteristics of the order data are analyzed to get the spatial-temporal factors that need to be considered when modeling.2.A prediction model based on hybrid deep learning method 3D-EDADF(3D Convolution and Encoder-Decoder Attention Demand Forecasting)is proposed to predict the inflow and outflow of travel demand in urban areas.The model uses 3D convolution and LSTM encoder-decoder to extract spatial-temporal features,and combines the attention mechanism to describe the difference between the inflow and outflow of travel demand.At the same time,the 3D-EDADF model carries out hybrid modeling of proximity dependence,daily dependence and periodic dependence,and weights and fuses their multi-dimensional features to obtain the final prediction result.Experimental verification is performed on two real data sets.The results show that compared with the baseline models,the 3D-EDADF model has better predictive performance,and can better explore and learn the spatial-temporal dependence and the spatial-temporal dependence of travel demand non-linear relationship.3.A 3D deconvolution-based deep learning regional travel demand prediction model RTDFF(Regional Travel Demand Forecasting Framework)is presented to predict the inflow and outflow of travel demand.The RTDFF model also uses fine-grained segmentation of time characteristics,including closeness dependence,daily dependence and periodic dependence.For each temporal characteristic,the model first uses a 3D convolutional layer to capture the long-term temporal dependence of spatial-temporal data and initially extract spatial features.Secondly,the 3D deconvolution layer is used to capture more detailed spatial features and quantify the importance of different regions,especially to pay more attention to hot spots,and improve the overall prediction performance of the model.Then the 2D residual unit layer is combined with to capture the farther spatial dependence and enrich the spatial features.Finally,the fusion layer is designed to integrate the multi-dimensional temporal and spatial characteristics and fine-grained temporal characteristics to improve the accuracy of the travel demand forecasting.A large number of experiments have been carried out using real travel demand data sets.The experimental results show that the RTDFF model has better predictive performance than the baseline models,which verifies the effectiveness of the RTDFF model.