基于 LSTM 神经网络的智能汽车轨迹跟踪混合控制策略研究

针对车辆轨迹跟踪过程中实时迭代求解隐式模型预测控制(model predictive control, MPC)计算效率低且实时性较差以及显式MPC受限于预计算的问题，文章提出基于长短期记忆(long short-term memory, LSTM)神经网络并结合MPC与线性二次型调节器(linear quadratic regulator, LQR)的混合控制策略。通过设置控制器的动态切换条件，在低速或状态误差较小时，采用LQR进行反馈控制，以快速响应并降低计算成本；在高速或状态误差较大时，采用MPC控制提供高精度轨迹跟踪，同时利用LSTM神经网络在线学习MPC控制行为，逐步逼近其控制效果，当LSTM学习误差满足阈值后，系统切换至LSTM结合LQR控制模式，LSTM生成基础控制信号，而LQR负责实时反馈调整以补偿动态环境中的扰动或误差。仿真实验结果表明，该混合控制策略在提高跟踪控制精度的同时能显著提升计算效率。

In order to solve the problem of low computational efficiency and poor real-time performance of iterative solution of implicit model predictive control (MPC) in the process of vehicle trajectory tracking, as well as the limitation of explicit MPC constrained by pre-computation, a hybrid control strategy based on long short-term memory (LSTM) neural network combined with MPC and linear quadratic regulator (LQR) is proposed. By setting the dynamic switching condition of the controller, LQR is used for feedback control at low speed or when the state error is small, so as to respond quickly and reduce the computational cost. At high speed or when the state error is large, MPC is used to provide high-precision trajectory tracking, and LSTM neural network is used to learn MPC control behavior online to gradually approximate its control effect. When the LSTM learning error meets the threshold, the system switches to the control mode of LSTM combined with LQR, LSTM generates the basic control signal, and LQR is responsible for real-time feedback adjustment to compensate for the disturbance or error in the dynamic environment. The simulation results show that the hybrid control strategy can improve the tracking control precision and computational efficiency significantly.