Data-enabled Predictive Control for Empty Vehicle Rebalancing

A critical operational challenge in Mobility-on-demand systems is the problem of imbalance between vehicle supply and passenger demand. However, conventional model-based methods require accurate parametric system models with complex nonlinear dynamics that are non-trivial to build or identify. In this paper, we implement a novel data-enabled predictive control algorithm for empty vehicle rebalancing (DeePC-VR) to instruct the repositioning policy between regions. Constructed by collected historical data from the considered unknown system, a non-parametric representation is used to predict future behavior and obtain optimal control actions, circumventing the costly system modeling process. The effectiveness of the proposed method is verified by an agent-based simulator modeling the real road network of Shenzhen, China. The proposed methods can serve more passengers with less waiting time compared to other policies, improving system efficiency and quality of service.