An Adaptive Model Predictive Control System for Virtual Coupling in Metros

Virtual coupling (VC) is an emerging concept and hot research topic in railways, especially for metro systems. Several unit trains in VC drive with a desired minimum distance, and they, as a whole, are regarded as a single train. In this work, a distributed adaptive model predictive control (AMPC) system is proposed to coordinate the driving of each unit train in VC. To obtain the accurate parameters of train dynamics model in a time varying environment, an estimator of the train dynamics model is designed for each AMPC controller. A variable step descent algorithm along the negative gradient direction is adopted for each estimator, which steers the estimated values of the parameters to real ones. Simulations are conducted and the results are compared with both nominal model predictive control system and AMPC system with fixed steps in the literature. Our proposed AMPC system with variable step (AMPCVS) has better performances than other two systems. Results indicate that there is an improvement of the proposed AMPC system with variable steps system when compared with other two existed systems. A running process of VC in a whole inter-station is also simulated here. Experimental results show that the trains track the desired objective well.

Energy-Efficient Data Collection Using Autonomous Underwater Glider: A Reinforcement Learning Formulation

The autonomous underwater glider has attracted enormous interest for underwater activities, especially in long-term and large-scale underwater data collection. In this paper, we focus on the application of gliders gathering data from underwater sensor networks over underwater acoustic channels. However, this application suffers from a rapidly time-varying environment and limited energy. To optimize the performance of data collection and maximize the network lifetime, we propose a distributed, energy-efficient sensor scheduling algorithm based on the multi-armed bandit formulation. Besides, we design an indexable threshold policy to tradeoff between the data quality and the collection delay. Moreover, to reduce the computational complexity, we divide the proposed algorithm into off-line computation and on-line scheduling parts. Simulation results indicate that the proposed policy significantly improves the performance of the data collection and reduces the energy consumption. They prove the effectiveness of the threshold, which could reduce the collection delay by at least 10% while guaranteeing the data quality.