Real-time Forecasting of Dockless Scooter-Sharing Demand: A Context-Aware Spatio-Temporal Multi-Graph Convolutional Network Approach

Abstract

Real-time demand forecasting for shared micromobility can greatly enhance its potential benefits and mitigate its adverse effects on urban mobility. The deep learning models provide researchers powerful tools to deal with the real-time dockless scooter-sharing demand prediction problem, but existing studies have not fully incorporated the features that are highly associated with the demand, such as weather conditions, demographic characteristics, and transportation supply. This paper proposes a novel deep learning model named Context-Aware Spatio-Temporal Multi-Graph Convolutional Network (CA-STMGCN) to forecast the real-time spatiotemporal dockless scooter-sharing demand. The proposed model applies a graph convolutional network (GCN) component that uses spatial adjacency graph, functional similarity graph, demographic similarity graph, and transportation supply similarity graph as input to extract spatial dependency and attach it to historical demand data. Then, we use a gated recurrent unit component to process the output of GCN and weather condition data to capture temporal dependency. A fully connected neural network layer is used to generate the final prediction. The proposed model is evaluated using the real-world dockless scooter-sharing demand data in Washington, D.C. The results show that CA-STMGCN significantly outperforms all the selected benchmark models, and the most important model component is the weather information. The proposed model can help the operators develop optimal vehicle rebalancing schemes and guide cities to regulate the dockless scooter-sharing usage.

Optimizing Shared E-Scooter Operations Under Demand Uncertainty: A Framework Integrating Machine Learning and Optimization Techniques

“…Similarly, to deal with demand sparsity, a masked fully convolutional network (i.e., fully convolutional network guided by a mask model or region of interest) was developed to focus on only the active cells in Calgary, Canada [31]. Additionally, Xu et al [32] used spatiotemporal multi-graph transformer-a graph convolutional network based on adjacency, functional similarity, demographic similarity, and transportation supply similarity graphs-to predict the hourly shared e-scooter demand in Austin, TX, and Washington, District of Columbia (DC). Furthermore, a long short-term memory (LSTM) was used to forecast the hourly demand of shared e-scooters in the Seocho and Gangnam districts of Seoul, South Korea [33].…”