Carsharing Demand Forecast And Intelligent Relocation Based On Rental Data

With the use of mobile internet and the development of the sharing economy under the background of "Carbon Peaking and Carbon Neutrality Goals",efficient,convenient and green carsharing services are becoming increasingly popular among travelers,and the large number of users also brings new challenges to the operation of carsharing system.In this paper,we forecast the demand of each station for station-based one-way carsharing system,and study the supply-demand balance and vehicle relocation at the stations based on the results of forecast.To address the station-based rental and return carsharing demand forecast problem,a station-embedding-based hybrid neural network(SEHNN)model is proposed to predict the number of rental and return vehicles at each station of the carsharing system.The model uses a variational graph auto-encoder(VGAE)module based on graph neural networks and long short-term memory neural network(LSTM)module to aggregate the spatial and temporal information.The VGAE performs station feature extraction and spatial latent embedding by combining fixed features such as the location,Points of Interest(POI)and temporal features such as weather and order interactions,and using graph convolutional neural networks(GCN)to integrate and normalize the embedding of these features based on the construction of order and geographic networks for the whole carsharing system;the LSTM is used to capture the temporal information of each station covering the whole network.The embedding represents the spatial and temporal information of each station and completes the prediction of rental and return demand at each station.In this paper,the model is validated with actual data from a carsharing company in Lanzhou,Gansu Province.The results show that compared with the traditional Elastic Net,ARIMA,LSTM,and Conv LSTM,the mean absolute error of the model for hourly rental/return prediction is reduced by 56.51%,47.18%,38.69%,and 38.52%,respectively,and the root mean square error is reduced by 42.68%,33.55%,25.68%,and 18.53%,respectively,and on this basis,the prediction results of the main stations and sub carsharing system from the carsharing network are compared,and it can be found that good prediction accuracy is achieved.To address the problem of intelligent relocation of station-based one-way car sharing system,we construct a reinforcement learning dynamic intelligent relocation framework based on user rental and return incentives under a fixed budget cost.The vehicle supply and budget of the current timestep in the carsharing system and demand forecast and incentive expenditure of the previous timestep are considered as states,the real-time incentive expenditure for each station is the decision or action,and the reward is the amount of unmet potential demand service that can be reduced.A Markov decision process based on the resource reduction policy and unknown future situation is constructed,and the reinforcement learning model is trained and solved using the PPO2algorithm(Proximal Policy Optimization).The effectiveness of the proposed model was evaluated by simulating the operation and order data of the carsharing system.Compared with the no-agent relocation method,the average service rate increased from24.1% to 50.8% in the state of tight vehicle resources and unbalanced regional demand,while the average service rate increased from 66.2% to 89.7% in the state of balanced regional demand.In the actual scene,the service rate increased by approximately 25%and hit 93.4%.In addition,when comparing the scenarios without vehicle demand forecasts and return incentives,the service rate increases for a given budget cost in combination with demand forecasts,while in consideration of return incentives,the service rate decreases when the budget cost is not sufficient but increases when the budget cost is sufficient.This paper contains 39 figures,6 tables and 92 references.