Test Cycle Simulation of Bench Based on Deep Reinforcement Learning

Abstract

<div>Test cycle simulation is an essential part of the vehicle-in-the-loop test, and the deep reinforcement learning algorithm model is able to accurately control the drastic change of speed during the simulated vehicle driving process. In order to conduct a simulated cycle test of the vehicle, a vehicle model including driver, battery, motor, transmission system, and vehicle dynamics is established in MATLAB/Simulink. Additionally, a bench load simulation system based on the speed-tracking algorithm of the forward model is established. Taking the driver model action as input and the vehicle gas/brake pedal opening as the action space, the deep deterministic policy gradient (DDPG) algorithm is used to update the entire model. This process yields the dynamic response of the output end of the bench model, ultimately producing the optimal intelligent driver model to simulate the vehicle’s completion of the World Light Vehicle Test Cycle (WLTC) on the bench. The results indicate that the algorithm exhibits good convergence in the simulation, throughout the WLTC simulation, the driver always kept the vehicle speed error within 1 km/h, and the response time is less than 0.5 s under the vehicle’s starting condition. In comparison to the PID control algorithm and the model predictive control (MPC) algorithm, it demonstrates smaller speed error and response time, ensuring accuracy, high efficiency, and safety during the indoor vehicle-in-the-loop test.</div>