Research On Demand Prediction And Dynamic Scheduling Of Shared Bicycles Based On Deep Learning

With the booming development of the sharing economy and low-carbon economy,shared bicycles are popular among the public for their low cost,stake-free parking and convenience.However,with the massive launch of shared bicycles,the phenomenon of waste bicycles piling up into "rubbish mountains" has emerged in various large and medium-sized cities,which not only affects the cityscape,but also causes waste of resources.In the long run,this will affect the development of the bicycle sharing industry.This requires that a generalized strategy of governance be put in place that is not just a "one-size-fits-all" approach to the parking of shared bicycles,but also a fine-tuned management that uses information technology and intelligent means to achieve a dynamic balance between the supply of and demand for shared bicycles.To ensure that parking spots are easily accessible and that parked bicycles do not impede traffic,parking rules need to be tailored to local conditions and fine-grained.As a representative method in the era of data intelligence,deep learning provides a powerful technical support for accurate demand prediction and timely dynamic scheduling of shared bicycles.Therefore,this paper takes demand prediction and dynamic scheduling as the grip,and combines relevant methods of deep learning to find feasible solutions for the robust operation of the bike-sharing model,which mainly includes the following three aspects of work.(1)Analyze supply and demand characteristics and perform regional clustering.This paper combines historical travel records to analyze the demand and supply characteristics of shared bicycles.The results show that at the demand level,the bicycle sharing system is highly complex and dynamic,with a non-linear and non-Euclidean correlation between rental and return demand in terms of spatial and temporal distribution;while at the supply level,the dynamic dispatch of shared bicycles is characterized by long-term benefits and process uncertainty.Based on the supply and demand characteristics of shared bicycles,this paper merges individual stations into regions based on internal equilibrium and interdependence,so as to facilitate further demand forecasting and dynamic dispatching.(2)Demand forecasting for shared bicycles based on graph neural networks.This paper combines graph neural networks with recurrent neural networks and designs a Spatio-Temporal Graph Convolution Neural Network(STGCN)to extract spatio-temporal features for the demand of shared bicycle rentals and returns,in order to improve the accuracy of demand forecasting.For spatial features,this paper adopts a multi-graph fusion approach,taking into account geographical proximity,traffic connectivity and functional similarity;for temporal features,this paper adopts a gating mechanism,thereby adjusting the weight of historical records at different time periods on the prediction results.(3)Dynamic scheduling of shared bicycles based on deep reinforcement learning.From the perspective of inventory path problem,this paper transforms the decision of loading and unloading quantity of dispatched vehicles and the decision of station access order in the dynamic dispatching of shared bicycles into an inventory optimization problem and a path optimization problem respectively,and builds the corresponding mathematical model based on Markov decision process(MDP).An innovative deep reinforcement learning method(DQN)is introduced to learn the state-action value functions involved in the dynamic inventory path problem of shared bicycles using its powerful function fitting capability,and to dynamically learn the combinations of decisions that should be taken by transport vehicles in the face of immobile environmental states through the time-series difference method.