Data-Driven LSTM Model and Predictive Control for Vehicle Lateral Motion

Abstract

This article proposes a methodology for designing Model Predictive Control (MPC) systems utilizing data-driven neural network model. The performance of MPC systems is highly dependent on the precision of the underlying model. While linearizing a vehicle’s dynamic model is feasible under certain conditions, accurately capturing the vehicle’s nonlinear dynamics, such as cornering stiffness and the interaction between tires and road surface, remains challenging. To address these nonlinear dynamics, this study proposes the estimation of a second-order lateral dynamic vehicle model via Long Short Term Memory (LSTM) model. The data-driven model utilizing LSTM has demonstrated better accuracy in simulating lateral vehicle motion when compared to conventional linearized dynamics model that employs nominal parameters. Ultimately, this data-driven LSTM model was incorporated into an MPC framework. This LSTM model-based MPC revealed enhanced performance in tracking accuracy over traditional MPC systems that utilize linear models for vehicle lateral dynamics.