Enhancing Ride Comfort and Reducing Energy Consumption of Ultra-High-Speed Elevators Using Deep Reinforcement Learning With Multivariate State Observation Reward Mechanism

Horizontal vibration of cars and high energy consumption of active controllers are two major problems in ultra-high-speed elevators, which will seriously affect ride comfort and cost. To address these issues, a novel active control method with deep reinforcement learning (DRL) based on a multivariate state observation (MSO) reward mechanism was proposed in this study. An eight-degree-of-freedom nonlinear dynamics model for horizontal vibration of cars in ultra-high-speed elevators was first established with precise simulation of the random guide rail excitation and the lateral aerodynamic forces. Subsequently, an MSO reward mechanism integrating multi-scale information such as vibration velocity, deflection angular velocity, and control force, was developed and added to the Softmax deep double deterministic policy gradients (SD3) algorithm. Finally, active control strategies for horizontal vibration were trained on the dynamics model using the SD3-MSO algorithm and three other DRL algorithms including deep deterministic policy gradient (DDPG), Twin delayed DDPG, and SD3. The results indicate that the active control using SD3-MSO algorithm could significantly reduce the horizontal vibration acceleration and deflection angular acceleration of the car. Vibration reduction performance is further improved by utilizing a smaller active control force. The root mean square (RMS) value of the active control force decreases by over 26% compared to the traditional fuzzy logic controller, demonstrating its superior energy-saving capacity.