Research On Shared Bike Demand Prediction And Scheduling Method Based On Deep Learning

After years of development,shared bicycles have become a highly intelligent mode of transportation in urban public transportation,providing green and convenient services for urban residents.However,with the increasing scale of shared bicycles,there have also been problems such as "disorderly parking" on urban streets,the occupation of urban green belts,users’ uncivilized use of vehicles,and damage to public facilities.In order to solve these problems and provide better travel services for users,sharing bicycle operation management companies have solved many site planning,delivery,and scheduling issues based on practical experience and methods over the years of operation.However,with the intellectualization of data and refinement of management,improving the efficiency of related work becomes increasingly important,requiring accurate vehicle delivery for large-scale sites,Timely dispatch the imbalance between supply and demand of vehicles at each station.Therefore,this article provides a new method and idea for predicting and scheduling shared bicycles using deep learning and intelligent algorithms.The main work is as follows:The spatiotemporal characteristics and important influencing factors of shared bicycle ridership were analyzed.Firstly,based on the shared bicycle system data of Citibike,New York City,USA,and the weather data of the city throughout the year,we found the 3141 station with the highest traffic volume,and counted the 8784 hours of riding volume and weather indicator data of the station throughout the year;The correlation between the number of shared bicycle rides and time factors such as months,days,and hours within a week,as well as weather factors such as temperature,weather type,and wind speed was studied.A prediction algorithm for shared single vehicle stations based on SSA-CNN-LSTM is proposed,which provides a method for calculating vehicle delivery volume for shared single vehicle operating companies.Firstly,the network architecture of CNN-LSTM was studied,using convolutional neural networks(CNN)and multiple long-term and short-term memory neural networks(LSTMs)to extract spatial features and learn time series information.The sparrow search algorithm was used to optimize the number of neurons in the LSTM layer,the size of convolutional cores,the number of convolutional layers,the number of neurons in the fully connected layer,and the learning rate in the model.The entire network structure diagram extracts spatial features from input data through CNN,and then sends them to LSTM to learn time series information.The sparrow search algorithm optimizes model parameters through search and update in each iteration.Finally,prediction experiments were conducted on data from two time dimensions of shared bicycle(three months and the whole year),and the results showed that SSA-CNN-LSTM was better than other benchmark models in effect fitting.The three evaluation indicators,RMSE,MAPE,and MAE,were significantly optimized.In the prediction experiments on January,April,August,and the whole year data,MAPE decreased by 10.4 percentage points,0.97 percentage points,6.78 percentage points,and 1.78 percentage points,respectively,RMSE and MAE also have some degree of optimization.A shared single vehicle scheduling algorithm based on adaptive large domain search under multi site hard time windows is proposed,providing a low-cost scheduling path for adjusting the imbalance in the number of vehicles between sites.Firstly,the issues of vehicle scheduling and route selection are analyzed.Based on the analysis results of the number of stations with the largest demand,3141 station is selected as the scheduling center,and the nearest 30 shared bicycle stations are selected as the scheduling demand stations.The longitude and latitude between the scheduling center and the station are converted into rectangular coordinates through Arcgis,and the distance between each station is calculated.Then,with the objective function of minimizing the total scheduling cost,a single vehicle regional scheduling optimization model with a hard time window is established.Taking the peak hour traffic on the evening of August 1st as the actual scheduling amount,the adaptive large domain search algorithm(ALNS)is used to solve the vehicle routing problem with a time window.Finally,a scheme with the lowest total scheduling cost was obtained: 12 scheduling vehicles were used to arrive at each station along the route at the specified time,meeting the actual needs of each station,and verifying the effectiveness of the algorithm for the shared single vehicle regional scheduling problem.