Research On Distribution Of Bicycle Stations And Prediction Of Bicycle Demand

Transportation is the most important point of society,it speeds up the efficiency of information communication and material exchange.There are multiple forms of transportation,such as walking,bicycling,public city bus,driving,etc.Riding,a trip mode which is approved by people because of its convenience and environmental power.Moreover,a lot of bicycle stations have been build up in many cities all over the world.However,there is a common issue when talking about bicycle rental,the imbalance fleet size between different stations:some stations have more bikes than the people need,while others have fewer bikes than people need.Therefore,to keep the lower price of shared bicycles,intentionally adjusting the deployment of bicycles in different stations in real-time is a non-trivial problem.In this scenario,traffic demand forecast is essentially required,and which is also one important part of ITS.In this thesis,we propose a novel demand forecast method for bicycle stations based on deep learning,which introduces a rational method of station divisions,and successfully increases the accuracy of demand forecast.More specifically.(1)We propose a novel station division method based on genetic algorithm.In this paper,the distance between the stations and the order exchange frequency are used as the basis for cluster division,which are also the optimization targets in the genetic algorithm.In the individual coding stage,we use edge-based encoding method to encode each possible cluster division plan,and the edge-based decoding method to decode the individual encoding quickly,which speeds up calculation of individual fitness.As for gene crossover and gene variation,the edge-based redivision method is used to enhance the diversity of individual genotypes.Finally,we analyze the rationality of division plan by measuring all trip demand point around the whole city.(2)A method of bicycle demand forecasting based on Ex-MGCN model is proposed.There are four various spatial relationship need to analyze: road network spatial relationship,bike lane spatial relationship,POI similarity spatial relationship,trip habit spatial relationship.In the temporal analysis,we use Long Short-Term Memory model to make effective forecast of bike demand according to dynamics of features based on the extraction of multiple spatial feature.Our experiment shows that,the proposed demand forecast method of bicycle stations via deep learning in this thesis is accurate and applicable in real-life scenarios.