Demand Prediction Of Dockless Shared Bikes Based On Deep Learning

Dockless bike-sharing systems play an important role in urban green traffic as a means of transportation that reduces greenhouse gas generation and provides flexible mobility.Bicycle demand is spatio-temporal imbalanced,it is easy to present unbalanced distribution phenomenon because of the lack of lock pile limitation.Therefore,it is an important part of the bicycle rebalancing task to excavate the travel characteristics of shared bikes in target cities and reasonably predict the bicycle demand.This paper integrates historical cycling data and POI data to establish demand prediction models from the aspect of the site level and network level respectively,providing data support for the rational placement of bikes.Firstly,analyze the characteristics of travel demand of users of shared bikes.On the basis of pre-processing the cycling data and POI data of Beijing Mobike in May2017,data mining method and spatial visualization tool are used to analyze the travel characteristics of bike users from the aspects of time,space and external conditions,and then determine the influencing factors of bike travel considered in this demand prediction study.The results show that there are significant differences in the time distribution of users’ riding demands on working days,holidays and different types of areas.The riding order is high around traffic,office and residential areas,and the riding distance is mainly concentrated within 1km.The weather has a significant impact on cycling demands.Secondly,establish a site-level demand prediction model based on LSTM.Combining cycling data,using the Geohash zoning generated virtual site,considering the date attribute and good weather condition,using LSTM network to extract time features and forecast demand a single area.Adjusting parameters to optimize prediction accuracy,then compared the prediction results with the traditional time series prediction model and the prediction effect without distinguishing date attributes.The results show that the 15 min time slice can achieve better prediction accuracy,the error is RMSE=9.79,MAE=7.42,which is better than the traditional time series prediction method,and the prediction accuracy can be improved when distinguishing date attributes.Finally,establish a network level demand prediction model based on GAT-LSTM.In order to improve the ability of spatial information capture,combining cycling data and POI data,introduce the graph attention neural network based on the site level demand forecasting model.Defining connected graph,functionally similar graph,interaction graph,and associated graph Sparse matrix to reflect spatial relationships,using multi attention to aggregate space information,and then the extracted characteristics of space time sequence is input into the LSTM for prediction.,fusing the single-graph prediction results.Analyze the influence of hyperparameters on the results and compare the fusion results with the single-graph prediction results,LSTM prediction results and GCN-LSTM prediction results respectively.The results show that multi-attention can improve the prediction accuracy and convergence speed;the prediction error was RMSE=3.40 and MAE=1.35,which were better than LSTM and GCN-LSTM;the result of multi-graph forecast is better than that of single graph,verifying the effectiveness of multi-graph attention mechanism and model.