Transformer-Based Prediction Model for Lateral Vehicle Dynamics

This paper investigates the suitability of a Transformer model as a prediction model in predictive control systems. Training data for the Transformer model was generated using a dynamic bicycle model. To account for measurement and process uncertainties, the outputs of the bicycle model were perturbed with additive Gaussian white noise. Subsequently, a compact Transformer model was trained for multi-step state forecasting. The prediction accuracy could be improved by applying a model that can natively capture temporal dependencies. Hence, the primary objective of this paper is to analyze the capability of transformer structures in replacing classical machine learning solutions for handling sequential data. Performance was evaluated by benchmarking the transformer model against the reference model. In all tested scenarios, the accuracy of the Transformer exceeded the preliminary expectations. Although further improvements can be implemented to enhance the performance of the model, and closed-loop embedding has not yet been performed, the evidence indicates that the Transformer can function as a predictor model.