Research On Taxi Travel Demand Forecasting Model Based On Regions

Taxi demand forecasting is a very important module in the construction of smart cities.For cities,an accurate taxi demand prediction system can not only help taxi companies increase efficiency,but also help reduce traffic jams and improve the urban air quality.In view of the outstanding performance of deep learning technology in learning the complex features and correlations of large data,a growing number of researchers have migrated deep learning to traffic data prediction.Most of the existing deep learning models choose to directly select all the input data or part of the data;this thesis,however,proposes a deep learning prediction model that incorporates clustering algorithms and correlation analysis into the selection process of input data,which is more comprehensive and accurate.In the prediction stage,this thesis put forward a Multi-Time Resolution Hierarchical Attention-Based Recurrent Highway Network(MTR-HRHN)model by expanding the existing sequence prediction model.It integrates the extraction of spatiotemporal features of exogenous data and spatiotemporal modeling of target data into a single framework.Furthermore,by introducing a layered attention mechanism,the model can adaptively select relevant exogenous features at different semantic levels.Finally,this model uses multiple time resolutions(such as every hour or every day),which expands the observable time pattern compared to a single time resolution situation.In the experimental part,the recorded data set of yellow taxi running in New York City from January to March 2019 is taken as the research object.The thesis conducts demand forecasting experiments in some areas in high demand with the model of this thesis and multiple classic forecasting models.The results show that the model in this thesis has better prediction in regional demand prediction,which not only improves the prediction accuracy by 47.69% on average compared to non-deep learning models,but also improves the prediction accuracy by 27.08% on average compared to other deep learning models.