Flow Prediction And Its Applications In Dockless Bike Sharing System

With the fast development of smartphones and the well-established internet-of-things(Io T)technology,the dockless bike-sharing system(DLBS)has proliferated domestically and abroad.In DLBS,the Io T lock equipped on the bike can be controlled remotely by the system on the cloud and keeps working for several months with a solar battery.The Io T lock frees the bike from the dock and increases the user experience by conveniently renting and returning a bike with a smartphone application.However,this new feature also brings a sequence of potential problems to the users and the operators.The system should inform users whether their potential destination is a parking-forbidden area from a user’s perspective.Also,in rush hours,the system should inform the user whether there is an upcoming bike when no bikes are available nearby to increase the user’s experience.From an operator’s perspective: they should plan many recommended parking places,estimate the number of bikes to be released accurately,and prepare for the hot-spot places stacking with bikes in the new city.Further,since the system is more imbalanced than the docked system,operators should pay more cost rebalancing the system.All these applications require a precise flow prediction both in macro-view and micro-view.For example,’informing parking-forbidden area’ and ’trip advice in rush hour’ applications are based on an accurate destination prediction of a single trip.Besides,’selecting recommended parking location’ and ’crowd-sourced rebalancing’ applications require an accurate inter-block bike flow prediction.However,the absence of docks in DLBS leads to multiple challenges when predicting the destination or flow:1.The selection of virtual stations will largely influence the model’s accuracy.The existing method of generating virtual stations is mostly based on the grid or start/end GPS clustering.However,the grid method mismatch with the functional block will result in the inaccuracy of the pre-diction.Moreover,the clustering-based method requires a large amount of data that cannot be applied in new cities.2.Predicting low-frequent user behavior is not trivial.As DLBS’s service area expands and provides convenience for the low-frequent users,the system’s data is sparser.Over half of the users’ trip records are no more than five times.Existing methods like matrix factor or auto-encoder require many parameters and are thus easy to over-fit.3.Predictions of a new city face the cold-start problem.There is no operational data to model when predicting a new city’s flow,and the knowledge transfer via geographical information is challenging due to the topology difference between cities.Existing works predicting the inter-block flow still require a small amount of operational data,together with the data in the source city.4.Inter-block flow prediction is difficult.Existing inflow/outflow prediction based on the station’s surrounding geography does not have sufficient precision to support the station selection and lane planning.Nevertheless,bike flow between the start/end blocks is influenced by the whole city’s block network and is thus difficult to predict.Moreover,the differences in topologies between the cities further aggravate the challenge.We follow the clue of analysis-model-application to study the dockless bike-sharing system systematically with these chapters: introduction to the dockless bike-sharing system,multi-view data analysis,destination prediction of a single trip in a micro view,cold-start bike flow prediction in a macro view,applications from introduction to operation and conclusion of the work.1.Chapter 1 illustrates the process of the dockless bike-sharing system’s development,its primary components,and main workflow and summarizes the current works from the aspects of macro-micro and user-operator.This chapter then discusses the difficult topics in this area.Moreover,propose our train of thought on the research of these problems.2.Chapter 2 illustrates the multi-view relationships of the geographical information,the user’s behavior,and the distribution of trips in DLBS.The system’s data reveals the POIs’(Point of Interest)clustering into functional blocks in the geographical view.In the user’s view,it reveals the power-law distribution of the riding frequency of users,and those high-frequent users’ behavior has an obvious pattern.The bike’s view reveals the unbalanced usage and short service life.The trip’s view reveals the peak flow temporal character in rush hour and the hot-spot spatial character of the bike flow.Moreover,the cities’ comparison view reveals the common ground and differences between the flows in those cities.3.Chapter 3 focuses on the problem of GPS level destination prediction in a micro view.This model uses POI as virtual docks and converts the start/end GPS pair into a set of virtual trips between POIs near those GPS based on the user’s tolerance of walking distance.The model acquires the hidden Markov transmitting matrix and finds the maximum likelihood estimated GPS based on the walking distance tolerance based on the historical records.To tackle the sparsity when deriving the transmitting matrix,the model uses a non-parameter feature extractor based on the fusion of prior,post,and compose probability which can prevent the overfit problem results from a large number of parameters.Comparisons with the model under various super-parameters with baseline models are illustrated to evaluate the effectiveness.4.Chapter4 focuses on the problem of cold-start inter-block flow prediction in a macro view and uses geographical information to transfer the knowledge between cities.This model divides the source/target city into functional blocks and blocks’ networks by the geographical road network.It regards the POI within a block as a fully connected graph with attributes,then derives embedding the block’s feature with an auto-encoder graph neural network.When predicting the inter-block flow,the model embeds the whole network of the city’s blocks into an “equivalent block of double port network”and transfers the knowledge between the geographical character and the bike flow between cities with different topologies.We further unify the problem of block feature embedding and bike flow prediction under a unified architecture and solve these problems via a relation graph network with a high order attention mechanism.Finally,we evaluate the model with a real-world data-set and show its effectiveness with super-parameter analysis.5.Chapter 5 discusses the applications in different stages from introduction to operation.This chapter classifies exciting parking management strategies and application scenarios in the estimating stage.In the infrastructure construction stage,this chapter considers two applications based on the inter-block flow prediction: planning the bike lane and locating the recommended parking places,modeling them in a mix-integer problem,and proving its effectiveness via simulations.In the efficiency promotion stage,this chapter considers two applications based on the micro and macro prediction: user’s trip advisor and crowd-source rebalancing proposes the algorithms and show their effectiveness with simulations.6.Chapter 6 concludes the work and analyzes the problems that require further investigation.